Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

A New Marketing Recommendation System Using a Hybrid Approach to Generate Smart Offers

A New Marketing Recommendation System Using a Hybrid Approach to Generate Smart Offers Applied Computer Systems ISSN 2255-8691 (online) ISSN 2255-8683 (print) December 2022, vol. 27, no. 2, pp. 149–158 https://doi.org/10.2478/acss-2022-0016 https://content.sciendo.com A New Marketing Recommendation System Using a Hybrid Approach to Generate Smart Offers 1* 2 Doae Mensouri , Abdellah Azmani 1,2 Intelligent Automation Laboratory, FST of Tangier Abdelmalek Essaadi University, Tetouan, Morocco Abstract – In order to increase sales, companies try their best to personalized global offers is one of the most efficient ways develop relevant offers that anticipate customer needs. One way to used. A global offer in marketing is defined as the combination achieve this is by leveraging artificial intelligence algorithms that of the main offer with a secondary offer. The main offer process data collected based on customer transactions, extract corresponds to a product or service that meets a basic customer insights and patterns from them, and then present them in a user- need. Whereas the secondary offering, also called an associated friendly way to human or artificial intelligence decision makers. offer, is the equivalent of the service or product offered along This study is based on a hybrid approach, it starts with an online marketplace dataset that contains many customers’ purchases and with the main one. Who is the customer? What will they buy? ends up with global personalized offers based on three different When will they buy it? And how much will they spend? These datasets. The first one, generated by a recommendation system, are classic questions that companies should answer to face the identifies for each customer a list of products they are most likely competition. Thus, the objective of the hybrid approach to buy. The second is generated with an Apriori algorithm. Apriori proposed in this study, which is an extension of a previous work is used as an associate rule mining technique to identify and map [2], is to generate global offers by identifying, for each frequent patterns based on support, confidence, and lift factors, and also to pull important rules between products. The third and customer, a list of the most likely to purchase products, the last one describes, for each customer, their purchase probability probability of purchase in the next few weeks, and the average in the next few weeks, based on the BG/NBD model and the transaction value. Furthermore, to make the most complete average of transactions using the Gamma-Gamma model, as well offers on the market, the associated products that are generally as the satisfaction based on the CLV and RFMTS models. By bought with the initial ones are determined as illustrated in combining all three datasets, specific and targeted promotion Fig. 1. strategies can be developed. Thus, the company is able to anticipate customer needs and generate the most appropriate offers for them while respecting their budget, with minimum II. CONTEXT operational costs and a high probability of purchase The success of any business mainly depends on an attractive transformation. marketing offer and an appropriate call to action. To facilitate Keywords – Apriori algorithm, BG/NBD Gamma-Gamma CLV the task of the customers, to attract more prospects and increase & RFMTS models, marketing recommendation system, smart company’s sales, it is very important to have a good business offer generation. strategy. In this section, the context and the utility of this work, which generate relevant and impactful marketing offers to I. INTRODUCTION attract the attention of target customers, are presented. Selling Nowadays, consumers are particularly concerned about the cannot be improvised; a client cannot be forced to buy. To sell, quality of the services and products offered. They are more customers’ needs should be first identified in order to offer a loyal to brands that make them feel pampered and privileged. product or service that meets their needs. To differentiate from However, the loyalty of customers is critical to the prosperity competitors, a global marketing offer is the key to success. A and success of a business [1]. Thus, companies are increasingly global offer combines the main offer that solves customer needs trying to anticipate customer needs and generate personalized with a secondary offer that contains associated items. product or service offers, which makes it possible to initiate, In this modern world, massive data can lead to information convert and build a relationship with customers. In order to overload problems for digital customers [3]. The consumer keep customers and increase their satisfaction, a marketing feels unable to find the suitable product among the multitude of strategy must extend far beyond conventional usage patterns, items offered on the online marketplace. To overcome this such as customer segmentation, addressing customers by first problem and help the consumer in decision-making, name in a promotional email, or showing relevant recommendation systems (RSs), proven to be a useful advertisements on a website. To provide a truly personalized technology [4], have been developed. These systems reduce the customer experience, companies use more innovative and number of options that the customer has to choose from by intelligent ways to customize the online experience; generating analysing the needs or behaviours. In general, recommendation Corresponding author’s e-mail: doae.mensouri@etu.uae.ac.ma ©2022 Doae Mensouri, Abdellah Azmani. This is an open access article licensed under the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0). Applied Computer Systems _________________________________________________________________________________________________2022/27 Fig. 1. Process of the offer generation approach. systems aim at helping customers make the right purchase presented in Section IV. Finally, Section V discusses the decisions [5]–[7] by analysing and aggregating suggestions conclusions and future research areas. from other customers [8]–[10]. Many websites, such as III. IMPLEMENTATION AliExpress, Netflix, YouTube, and Spotify, use recommendation-based algorithms [11]. The wide application Instead of identifying targeted customers for a specific offer, of those recommender systems, based on either the deep this study takes the opposite route and generates offers by learning approach or the machine learning approach, in the e- anticipating customer needs. As an effective tool to help commerce marketplace is of increasing interest to companies develop targeted marketing strategies, the approach researchers [12]–[21]. While some studies have focused on presented in this paper, as shown in Fig. 1, exploits artificial targeted offers, i.e., the selection of customers who will receive intelligence techniques deployed in parallel sections to create an offer consisting of one or more products during the campaign personalized offers by: [22], [23], this study generates personalized offers that best • Building a recommendation system by anticipating match the preferences of each customer. customer needs based on historical data, then generating Consumers are constantly bombarded, mainly by email [24], association rules to identify frequent patterns between with proposals for offers that often do not correspond to their items and to identify associated ones; profile or their needs. On the supplier’s side, this translates into • Identifying the probability of purchase in the next 4, 8, and additional costs since it is necessary to send several thousand 12 weeks; emails to be able to achieve poor results and into a questioning • Predicting the value of next purchase; of its image because a large number of people perceive its • Calculating customer’s profitability and satisfaction; proposals as an annoyance and a saturation of their email [25]. • Generating a global, personalized offer. It is to remedy such problems that the approach proposed in the A. About the Used Dataset context of this article is inscribed and which generates an irresistible and personalized offer that best matches the The dataset used in this article is called “Brazilian E- customer’s preferences, thus, on the one hand, providing value Commerce Public Dataset”, provided by Olist Store and and enticing customers to purchase and, on the other hand, accessible on the Kaggle website (https://www.kaggle.com). efficiently reducing the problem of information overload. The This dataset contains multiple purchases from 2016 to 2018 at delivered offers have a high transformation factor due to the various stores using the Olist service. In this study, only the combination of several methods such as recommendation Customers (99 441 rows), Orders (99 478 rows), Reviews systems, the Apriori algorithm, association rules, the BG/NBD (98 410 rows) and Payments (99 478 rows) sub-datasets are model, and the Gamma-Gamma model, as explained in the used since they contain the useful variables for this study following sections. approach. The three sub-datasets are filtered and prepared, then This article is organised as follows: Section II provides an merged by their respective primary keys into a single initial overview of the context related to marketing offer generation. dataset as shown in Table I. Section III presents the construction and evaluation of the In order to identify the probability of future purchases in the various models used. A discussion of the obtained results is next few weeks, the value of the next one, as well as customer Applied Computer Systems _________________________________________________________________________________________________2022/27 satisfaction, five variables, which are recency (R), frequency neighbours to recommend an article to a target customer [3], (F), monetary value (M), interpurchase time (T), and [27]. satisfaction (S), should be first calculated. To do so, the RFMTS Collaborative Filtering Model (CF) model [2] is used, Table II presents the dataset after generating Recommendation systems often use collaborative filtering. It the R, F, M, T, and S variables for each customer. builds on the idea that the best recommendations come from • R: represents the period between the current moment and other customers with similar taste. This means using similar the customer’s last purchase. A low recency rate indicates customer review history to predict how a customer will rate a that customers buy frequently. product. CF models can recommend a product to any user X • F: represents the number of transactions made by a based on the preferences of a like-minded user Y [28]. customer. A high frequency rate indicates a high level of Collaborative filtering has two subcategories: model-based and loyalty. memory-based. This study covers collective memory-based • M: represents the average amount of a specific customer’s filtering given that user-based (user-user) and item-based (item- purchases. A high monetary value implies that the to-item) filtering belong to this category. The difference customer brings high profits to the company. between the two mainly lies in the considerations when • T: represents the time between a customer’s two calculating the recommendation. successive purchases. • S: represents the total of a customer’s ratings divided by • User-based collaborative filtering: it focuses on the the number of reviews. “nearest neighbour” approach [29] to recommend items by looking at user ratings to find the “nearest neighbours” TABLE I with ratings similar to the targeted customer. In other INITIAL DATASET words, user-based filtering finds users with similar Order Customer Payment Review spending habits and provides recommendations based on Order ID purchase ID value score time their interests. 10/2/2017 4 • Item-based collaborative filtering: first introduced by e4…f7 10.11 000…1e2 10:56 Amazon in 1998 [30]. Contrary to user-based 10/2/2017 4 collaborative filters, item-based filtering finds patterns of 9c…3c 18.12 000…be3 10:56 similarity between different elements by considering the 10/2/2017 3 number of users who have purchased items A and B 11…62 25.59 000…064 10:56 simultaneously. If the correlation between item A and B is … … … … high enough, it can be assumed that the two items are 142.14 4 7/24/2018 similar to each other. As a result, customers who purchase 53…51 fff…684 20:41 item B will receive a recommendation of item A and vice 8/8/2018 1 versa. 47…5d 20.46 fff…30a 8:38 Fig. 2 shows the various collaborative filtering approaches TABLE II based on the user and the item. The dotted arrows on the left DATASET AFTER CALCULATING R, F, M, T AND S VALUES represent recommendations based on user similarity and Customer preference, and the dotted arrows on the right represent R F M T S ID recommendations based on similar items. 000…1e2 1428.77 2.10 120.66 432.67 1.20 Evaluation 000…be3 230.48 2.24 180.04 345.89 3.15 In machine learning, the accuracy of a model is evaluated by 000…064 428.96 1.63 80.23 115.66 4.22 mathematically taking the root mean square of the error that has … … … … … … occurred between the test and the predicted values [31]. This fff…684 149.52 1.96 71.99 109.71 4.64 error value is used as a defensive score to decrease the fff…30a 230.41 2.15 44.84 84.32 3.27 individual effects of individual measurement errors so that there are no significant assumptions that substantially distort the B. Recommendation Systems results [11]. The RMSE (Root Mean Square Error) is a method In recommender systems, two algorithms are mainly used: for determining the accuracy of a predictive target value model content-based filtering (CBF) and collaborative filtering (CF). [32]. It is the estimate of the average difference between the The CBF technique is primarily based on customer and product actual value and the estimated value, which is calculated as information. It recommends a product to a customer based on shown in (1). Where n is the number of estimates, yj represents the profile, feedback, and item information, etc. Its main the current estimate, and ŷj represents the average value of the restriction is the limited analysis of content as well as the estimates. The comparison of the user-based model and item- specialization [26]. CF recommends products to consumers based model results based on the RMSE value is represented based on their previous ratings to alleviate the above problems. Table III. This shows that the user-based model gave a better This technique is based on the preferences of the top k similar result, with a minimum RMSE error value. Applied Computer Systems _________________________________________________________________________________________________2022/27 Fig. 2. User-based and item-based collaborative filtering approaches. [45]. Rules, whose lift is greater than 1, are the practicable association rules, which implies that item A has a positive effect 1 2 RMSE YY− . (1) on the appearance of item B. The support for a set of items is ( ) ∑ jj j=1 the percentage of the dataset that contains this set of items [50], [51]. Let I = {i , i ,...,i } be a set of items and A – a set of some a b z The implementation of the user-based model on the dataset items in I. Equation (2) defines the support of A, where identifies the top 5 recommended items for each customer, Count (A) is the number of times item-set A appears in all based on customer’s preferences (as shown in Table IV). transactions and N is the number of transactions. TABLE III C. Generating Association Rules by Applying Apriori Algorithm USER-BASED AND THE ITEM-BASED MODELS RESULTS BASED ON THE RMSE Market basket analysis (MBA) is a data mining technique VALUE commonly used to identify which products tend to be purchased Model RMSE together [33], [34] and predict customer purchasing behaviour. user-based model 1.063122817661801 It discovers the association between the activities performed by item-based model 1.343641161111319 the customers [35]. Although the primary focus of MBA is on shopping carts and customers [36]–[38], it also includes fields such as highway-railroad safety [39], fraud detection [40], Count (A) mobile showroom [41], and others. MBA tries to identify Support (A) = . (2) 𝑁𝑁 purchase patterns in a large number of customer transactions. This is where association rules play an essential role by finding For rule A⇒B (antecedent →consequent), the Support, relations between different items and discovering frequent ones Confidence and Lift equations are the same presented in [52]: [34], [42]. Association rules do not capture individual preferences but instead look for relationships between different • The support of the A→B association rule is the sets of elements in each transaction. This is what makes them percentage of transactions that include both elements different from collaborative filtering, which is used in A and B. It is defined as shown in (3). recommendation systems [43]. To obtain the networks of {A,B} products, the Apriori algorithm, the most frequent and efficient Support (A → B) = . (3) 𝑁𝑁 algorithm for association rule generation [44], [45], is applied in this research to generate these rules. • Confidence measures the number of times items in B The Apriori algorithm was the first to determine Boolean appear in transactions involving A [53]. The association rules by extracting patterns from datasets [46]. It is Confidence is calculated as shown in (4). used in different fields, in medicine to discover frequent image patterns in mammogram images [47], in education to predict {A,B} Confidence (𝐴𝐴 → B) = . (4) Support (A) system of learning [48], in traffic accidents to identify the main causes and trends related with it [49]. The Apriori algorithm is • Lift is a probability that event B occurs independently a method to find recurring items in a transactional dataset and as a consequence of another event A [35]. to generate associated rules. It first determines frequent, unique Equation (5) defines the lift of the association rule items through transactions, then generates association rules A→B. [45]. The set of frequently occurring items corresponds to the set of frequently occurring functions measured by the parameter {A,B} Lift (A → B) = . (5) support. Confidence and lift parameters are used to measure the Support (B) association rules corresponding to implication rules (A→B) = Applied Computer Systems _________________________________________________________________________________________________2022/27 The traditional approach to association rules is to mine each customer transaction set produced meaningful association records for common item-sets with a minimum threshold of rules with high confidence (greater than 0.7) as shown in “support” and then use another minimum threshold of Table V. These results, which determine which products are “confidence” to generate rules for the collected common item- frequently bought together, allow companies to develop sets [54] (Fig. 3). Applying the priori association algorithm to customer loyalty strategies by generating global offers. Fig. 3. Association rule mining. The application of the BG/NBD model to this article dataset, TABLE IV including the RFT data generated previously, makes it possible TOP 5 RECOMMENDATIONS FOR EACH CUSTOMER to create the Frequency-Recency Matrix shown in Fig. 4. The Customer ID Top 5 recommended items visualization of this matrix, over four weeks (Fig. 4), shows the B00CGW74YU (Laptop), probability of future customer purchases at a given time. This B000Q8UAWY (Phone), probability increases if the number of purchases and the 000…1e2 B003TXZCYE (Perfume), B008DW95NA (T-shirt), duration since its last purchase are important. It should be noted B005N3TPJQ (Gaming mouse) that the probabilities of future purchases are quite low over the B000P6G7YW (Perfume), next four weeks and are limited to a few customers; this is B00000JX3C (Baby toy), explained by the pattern of sporadic purchases seen in the 000…be3 B000GC2DDE (Baby clothes), studied dataset. B001PB17YY (Backpack), B0022TQXWG (Diaper) Based on the previous model, predictions could be made for any chosen period. In the case of this study, predictions are B0002KR13M (Dress), B000VAIMF4 (Sunscreen), created for the first 4, 8, and 12 weeks, as these are a strong 000…064 B002ZYQE0Y (Perfume), short-term indicator for focusing a vendor’s efforts on their B0058SYDTI (Swimsuit), customers with the highest likelihood of future purchases. B004TLL730 (Makeup) Table VI shows the probability of purchase in the next 4, 8, and … … 12 weeks for each customer. TABLE V BG/NBD Validation ASSOCIATION RULES Model predictions cannot be used blindly, they must be Rule Antecedents Consequents Support Conf Lift validated with reality. For this purpose, the last three months of 0 e0c…a26 368…2db 0.37 0.79 1.68 results, when the numbers of transactions of each customer are … … … … … … already known, are compared with the BG/NBD model results. 125 389…0f4 0bc…cdd 0.40 0.84 1.68 As can be seen in Fig. 5, the lines are quite close to each 126 0bc…894 537…48c 0.39 0.80 1.70 other, which means that the model makes predictions quite 127 537…f73 0bc…48c 0.41 0.86 1.70 close to the margin of error. In all cases, the model is close on certain points to reality. Similarly, extending the duration of the … … … … … … test could improve the validation of the model. E. Predictive Calculation of the Future Purchase Value by D. Calculation of the Probability of a Future Purchase by Application of the Gamma-Gamma Model Application of the BG / NBD Model The probability of having each customer’s purchases forecast The BG/NBD model [55] is used in this study to estimate, for for the following months can be known, but not their number. each customer, the probability of carrying out transactions in a A customer who buys 20 times on average 5 USD is not the future period. The BG/NBD is a simplified substitutional model same as another who buys 2 times 400 USD. It is at this stage to the well-known Pareto/NBD model [56], which is based on that the Gamma-Gamma model initially proposed by Colombo Bayesian probability in a hierarchical way to make its and Jiang [60] is used to give an estimate of the profitability of estimates. These two models are particularly used in marketing each customer, using the monetary value (M) column research, especially in the analysis of the customer database. previously generated. They are rooted in the theory of customer lifetime value [57]– One of the requirements of the Gamma-Gamma model is to [59] that forecasts future customer behaviour based on only work with observations with a purchase frequency stochastic aspects of interaction behaviour, such as different from 0 (retain only customers with more than one departure [56]. purchase) [61]. Similarly, it must be ensured that there is no Applied Computer Systems _________________________________________________________________________________________________2022/27 TABLE VI correlation between the frequency (F) and the monetary PURCHASE PREDICTIONS OF NEXT 4, 8, AND 12 WEEKS value (M) of each consumer [62]. In practice, the Gamma- Gamma model must check whether the Pearson correlation [63] Expected Expected Expected Customer ID 4 weeks 8 weeks 12 weeks between the two vectors is close to 0 (Fig. 6). 000…1e2 0.02 0.04 0.05 One of the faculties of this model is to calculate the conditional expectation of the average profit per transaction for 000…be3 0.03 0.07 0.01 each customer as shown in Table VII. 000…064 0.02 0.04 0.07 … … … … F. Customer Profitability and Satisfaction fff…84 0.02 0.04 0.07 Customer Lifetime Value (CLV or CLTV) is a marketing metric [64] that, in a nutshell, reflects the profit that a certain ffff371b4d645b6ecea244b27531430a 0.02 0.01 0.01 customer will provide to a business over a given period of time. TABLE VII This makes it possible to get out of the subjectivism of the value of a customer, in order to quantify it in a more exploitable field CUSTOMER’S NEXT AVERAGE TRANSACTION VALUE linked to the profitability it brings to the company. The concept Customer ID Avg transaction of customer lifetime value is well accepted by researchers and 0000366f3b9a7992bf8c76cfdf3221e2 12 398.64 marketers [58] who agree that long-term customers are more 0000b849f77a49e4a4ce2b2a4ca5be3 4542.78 profitable. The reason commonly given for increasing the 0000f46a3911fa3c0805444483337064 6608.59 profitability of long-term customers is the increase in profit … … from loyal customers, which is determined by the price fffcf5a5ff07b0908bd4e2dbc735a684 344.3 premium they pay, the additional profit from sales through referrals, and the benefits in relation to cost savings due to the ffff371b4d645b6ecea244b27531430a 69.25 increased sales in their regard [65]. TABLE VIII NORMALIZED CLV AND S Customer unique id CLV S 0000366f3b9a7992bf8c76cfdf3221e2 0.928328 0.05 0000b849f77a49e4a4ce2b2a4ca5be3 0.870307 0.5375 0000f46a3911fa3c0805444483337064 0.662116 0.805 … … fffcf5a5ff07b0908bd4e2dbc735a684 0.098976 0.91 ffff371b4d645b6ecea244b27531430a 0.095563 0.5675 Customer satisfaction and trust are respectively the first and second important antecedents of customer loyalty [66]. After calculating the CLV for each customer, based on the RFMTS model [2], the CLV and satisfaction (S) values are normalized to a range between 0 and 1, using the well-known statistic normalization formula (6), as indicated in Table VIII. 𝑋𝑋𝑋𝑋 −min(𝑋𝑋 ) Fig. 4. Frequency recency matrix. 𝑋𝑋 norm = . (6) ( ) max 𝑋𝑋 −min(𝑋𝑋 ) Fig. 6. Frequency & Monterey correlation matrix. Fig. 5. Actual purchases in holdout period vs. predicted purchase. Applied Computer Systems _________________________________________________________________________________________________2022/27 IV. RESULTS AND DISCUSSION: HOW TO CREATE A GLOBAL algorithm, so generation is automatic which takes as input OFFER BASED ON DIFFERENT INFORMATION GENERATED information about customers and their purchases and outputs PREVIOUSLY generated offers. The global algorithm developed in this study makes it possible to implement parallel sections of the process Predicting future customer behaviour and generating a simultaneously to reduce processing time, and then the results dedicated global offer for each of them allows the company to of each section are combined to provide a personalized offer. consolidate its customer relationships, differentiate itself from its competitors, and increase its earnings. It is in this context V. VALIDATION that the study presented here takes place, with the aim of anticipating customer needs in order to generate a global offer In addition to the different evaluation factors concerning by following the steps illustrated in Fig. 1. The global offer is each technique used at different stages of this study, the offer made up of the main offer and associated offer for customers transformation rate is used in the end to validate and evaluate who are probably likely to accept the offer positively based on the proposed offer generation algorithm. The conversion rate is their needs. The main offer corresponds to a product or service a key factor in the financial success of a platform [67]. It is a that satisfies consumer’s basic needs. While associated offer is percentage metric that helps companies analyse the added to the main one with additional products or services that performance of their marketing strategies and can be calculated are generally purchased with the bought one. The global offer by dividing the number of sales by the total number of creation is divided into two steps. The first step is the main offer communicated offers and multiplying by 100 %, as shown creation; in this step a recommendation system is used to in (7). identify, for each customer, the top 5 products that meet the Number of sales customer needs (Table IV). The second step is the associated Conversion rate= · 100. (7) Number of communicated offers offer creation; in this step Apriori algorithm is used to determine associated products as shown in Table V. After To validate our approach, the data are split into two parts, determining the products of interest to a specific customer, as with the last 20 % used for the validation and 80 % used for well as the associated products, the third step takes place, which testing. The analysis of the 80 % of data has made it possible to aims to calculate the probability for a customer to transact in a determine for each customer the products of interest, the future period using the BG/NBD model. The probabilities of associated products, purchase time and purchase amount, and customer’s future purchases in the next 4, 8 and 12 weeks give the value of CLV and satisfaction. Let us take as an example a an idea of the ideal time to communicate to the customer a set of 65 customers identified by our model as being interested personalized offer. With the above, the probability of each in the product “laptop” and that their CLV and satisfaction customer purchase in the next few months is known but not its scores are close. Fig. 1 represents the personalized offer amount. This is where the role of the next step comes in, which generated for those 65 customers. The conversion rate value of consists in calculating the future purchase value by application this offer, calculated by (7), is 70 %, which is a high value that of the Gamma-Gamma model. Knowing the value of the next means an average of 45 customers out of 65 positively purchase helps choose the associated products, identified in interacted with the generated offer. Step 2, while respecting the customer’s budget, in other words, the top products, the total price of which does not exceed the VI. CONCLUSION client’s next purchase value, are chosen. Combining all the Recent advances in technology are impacting not just information generated in the above steps makes it possible to consumers, but merchants as well. On the one hand, today’s generate a global personalized offer with a high transform consumers are constantly changing, and digitization is the value. All customers are not the same, and so they should not major reason for this change in behaviour. On the other hand, be treated the same way. A loyal customer deserves eventually AI algorithms have given marketers the ability to process and more promotions and services added to the offer. Thus, to extract insights from collected customer transaction data and identify the value of a customer, it is necessary to calculate the thereby increase profitability with minimum operational cost, CLV value (calculated based on customer recency, frequency which is the goal of any type of business. Hence, this article has and monetary value) and satisfaction value (calculated based on represented a process of generating personalized offers as part customer review rating). By following the approach proposed of an overall marketing strategy aimed at anticipating customer in this study, a global offer that contains associated products needs and developing targeted and specific promotional and personalized services is generated for each customer. In strategies. First, a recommendation system is built based on addition to the probability of purchase, the global offer respects customer historical data, and then association rules are the customer’s budget as well. Fig. 7 represents an example of generated to identify frequent patterns between items. After a personalized offer generated based on the approach described determining what each customer may buy, the next step is to above for a specific customer, where “Laptop” is the main determine the best time and the value of the next purchase. customer’s need, and the remaining items are additional ones Finally, the profitability and satisfaction of each customer are with the additional option of checking and unchecking the items identified. proposed while adapting the price to make it attractive. All Due to the approach presented in this study, customer methods used in this study are integrated into a global overload problems are minimised since the customers receive Applied Computer Systems _________________________________________________________________________________________________2022/27 personalized offers about products that meet their needs. approach presented in this paper can be reinforced by taking Competition is getting tougher, and to differentiate themselves into consideration other parameters, in particular the probability from competitors and increase the conversion rate of offers, of purchase, the profile of the customer, as well as the budget, companies must forget about old marketing strategies and opt to plan the sending of the offer while respecting the customer’s for personalization. In addition to the customization of the schedule and personalizing the associated message according to offers, which is the objective of this study, choosing the right the profile (age, gender, residence, etc.). moment and the right sending channel to send the offer increases customer satisfaction. Thus, in future research, the Fig. 7. Global offer example. [6] N. Jing, T. Jiang, J. Du, and V. Sugumaran, “Personalized CKNOWLEDGMENT recommendation based on customer preference mining and sentiment The research has been supported by the Ministry of Higher assessment from a Chinese e-commerce website,” Electronic Commerce Research, vol. 18, no. 1, pp. 159–179, Nov. 2018. Education, Scientific Research and Innovation, the Digital https://doi.org/10.1007/s10660-017-9275-6 Development Agency (DDA) and the National Centre for [7] H. Zhang, L. Zhao, and S. Gupta, “The role of online product Scientific and Technical Research (CNRST) of Morocco recommendations on customer decision making and loyalty in social shopping communities,” International Journal of Information (Smart DLSP Project – AL KHAWARIZMI IA-PROGRAM). Management, vol. 38, no. 1, pp. 150–166, Feb. 2018. https://doi.org/10.1016/j.ijinfomgt.2017.07.006 REFERENCES [8] E. Frias-Martinez, S. Y. Chen, and X. Liu, “Evaluation of a personalized digital library based on cognitive styles: Adaptivity vs. adaptability,” [1] K. Matsuoka, “Effects of revenue management on perceived value, International Journal of Information Management, vol. 29, no. 1, pp. 48– customer satisfaction, and customer loyalty,” Journal of Business 56, Feb. 2009. https://doi.org/10.1016/j.ijinfomgt.2008.01.012 Research, vol. 148, pp. 131–148, Sep. 2022. [9] E. Frias-Martinez, G. Magoulas, S. Chen, and R. Macredie, “Automated https://doi.org/10.1016/j.jbusres.2022.04.052 user modeling for personalized digital libraries,” International Journal of [2] D. Mensouri, A. Azmani, and M. Azmani, “K-Means customers Information Management, vol. 26, no. 3, pp. 234–248, Jun. 2006. clustering by their RFMT and score satisfaction analysis,” International https://doi.org/10.1016/j.ijinfomgt.2006.02.006 Journal of Advanced Computer Science and Applications (IJACSA), [10] H.-N. Kim, A.-T. Ji, I. Ha, and G.-S. Jo, “Collaborative filtering based on vol. 13, no. 6, 2022, Art no. 6. collaborative tagging for enhancing the quality of recommendation,” https://doi.org/10.14569/IJACSA.2022.0130658 Electron. Commer. Res. Appl., vol.9, no. 1, pp. 73–83, Jan.–Feb. 2010. [3] P. K. Singh, E. Othman, R. Ahmed, A. Mahmood, H. Dhahri, and https://doi.org/10.1016/j.elerap.2009.08.004 P. Choudhury, “Optimized recommendations by user profiling using [11] P. Boström and M. Filipsson, “Comparison of user based and item based apriori algorithm,” Applied Soft Computing, vol. 106, Jul. 2021, Art. no. collaborative filtering recommendation services,” 2017. [Online]. 107272. https://doi.org/10.1016/j.asoc.2021.107272 Available: https://www.semanticscholar.org/paper/Comparison-of-User- [4] B. Tilahun, C. Awono, and B. Batchakui, “A survey of state-of-the-art: Based-and-Item-Based-Filtering-Bostr%C3%B6m- Deep learning methods on recommender system,” IJCA, vol. 162, no. 10, Filipsson/27984027b4fccd7323371129768c459ba16f8fbd. Accessed on: pp. 17–22, Mar. 2017. https://doi.org/10.5120/ijca2017913361 Mar. 04, 2022. [5] I. Y. Choi, M. G. Oh, J. K. Kim, and Y. U. Ryu, “Collaborative filtering [12] Y. Chen, Y. Lu, B. Wang, and Z. Pan, “How do product recommendations with facial expressions for online video recommendation,” International affect impulse buying? An empirical study on WeChat social commerce,” Journal of Information Management, vol. 36, no. 3, pp. 397–402, Jun. Information & Management, vol. 56, no. 2, pp. 236–248, Mar. 2019. 2016. https://doi.org/10.1016/j.ijinfomgt.2016.01.005 https://doi.org/10.1016/j.im.2018.09.002 Applied Computer Systems _________________________________________________________________________________________________2022/27 [13] T. Donkers, B. Loepp, and J. Ziegler, “Sequential user-based recurrent [30] G. D. Linden, J. A. Jacobi, and E. A. Benson, “Collaborative neural network recommendations,” in Proceedings of the Eleventh ACM recommendations using item-to-item similarity mappings,” Conference on Recommender Systems, New York, NY, USA, Aug. 2017, U.S. patent US6266649B1, 1998. [Online]. Available: pp. 152–160. https://doi.org/10.1145/3109859.3109877 https://patents.google.com/patent/US6266649/en. Accessed on: Mar. 04, [14] B. Hidasi and A. Karatzoglou, “Recurrent neural networks with top-k 2022. gains for session-based recommendations,” in Proceedings of the 27th [31] S. Jayalakshmi, N. Ganesh, R. Čep, and J. S. Murugan, “Movie ACM International Conference on Information and Knowledge recommender systems: Concepts, methods, challenges, and future Management, New York, NY, USA, Oct. 2018, pp. 843–852. directions,” Sensors, vol. 22, no. 13, Jan. 2022, Art. no. 13. https://doi.org/10.1145/3269206.3271761 https://doi.org/10.3390/s22134904 [15] I. Islek and S. G. Oguducu, “A hierarchical recommendation system for [32] Ç. Odabaşı, P. Dologlu, F. Gülmez, G. Kuşoğlu, and Ö. Çağlar, E-commerce using online user reviews,” Electronic Commerce Research “Investigation of the factors affecting reverse osmosis membrane and Applications, vol. 52, Mar.–Apr. 2022, Art. no. 101131. performance using machine-learning techniques,” Computers & https://doi.org/10.1016/j.elerap.2022.101131 Chemical Engineering, vol. 159, Mar. 2022, Art no. 107669. [16] Z. Li, H. Zhao, Q. Liu, Z. Huang, T. Mei, and E. Chen, “Learning from https://doi.org/10.1016/j.compchemeng.2022.107669 history and present: Next-item recommendation via discriminatively [33] P. Jirapatsil and N. Phumchusri, “Market basket analysis for fresh exploiting user behaviors,” in Proceedings of the 24th ACM SIGKDD products location improvement: A case study of E-commerce business International Conference on Knowledge Discovery & Data Mining, New warehouse,” in Proceedings of the 4th International Conference on York, NY, USA, Jul. 2018, pp. 1734–1743. Management Science and Industrial Engineering, New York, NY, USA, https://doi.org/10.1145/3219819.3220014 Apr. 2022, pp. 23–28. https://doi.org/10.1145/3535782.3535786 [17] N. Sachdeva and J. McAuley, “How useful are reviews for [34] A. Monteserin and M. G. Armentano, “Influence-based approach to recommendation? A critical review and potential improvements,” in market basket analysis,” Information Systems, vol. 78, pp. 214–224, Nov. Proceedings of the 43rd International ACM SIGIR Conference on 2018. https://doi.org/10.1016/j.is.2018.01.008 Research and Development in Information Retrieval, New York, NY, [35] S. Halim, T. Octavia, and C. Alianto, “Designing facility layout of an USA, Jul. 2020, pp. 1845–1848. amusement arcade using market basket analysis,” Procedia Computer https://doi.org/10.1145/3397271.3401281 Science, vol. 161, pp. 623–629, Jan. 2019. [18] B. Shao, X. Li, and G. Bian, “A survey of research hotspots and frontier https://doi.org/10.1016/j.procs.2019.11.165 trends of recommendation systems from the perspective of knowledge [36] Y. Chen, “Research on personalized recommendation algorithm based on graph,” Expert Systems with Applications, vol. 165, Mar. 2021, Art no. user preference in mobile e-commerce,” Information Systems and e- 113764. https://doi.org/10.1016/j.eswa.2020.113764 Business Management, vol. 18, pp. 837–850, Dec. 2020. [19] D. R. Stöckli and H. Khobzi, “Recommendation systems and convergence https://doi.org/10.1007/s10257-019-00401-2 of online reviews: The type of product network matters!” Decision [37] M. Gupta, S. Kochhar, P. Jain, and P. Nagrath, “Hybrid recommender Support Systems, vol. 142, Mar. 2021, Art. no. 113475. system using A-priori algorithm.” in Proceedings of International https://doi.org/10.1016/j.dss.2020.113475 Conference on Sustainable Computing in Science, Technology and [20] J. Zhang, M. Li, W. Liu, S. Lauria, and X. Liu, “Many-objective Management (SUSCOM), Jaipur, India, Feb. 24, 2019. optimization meets recommendation systems: A food recommendation https://doi.org/10.2139/ssrn.3349290 scenario,” Neurocomputing, vol. 503, pp. 109–117, Sep. 2022. [38] R. Katarya and O. P. Verma, “A collaborative recommender system https://doi.org/10.1016/j.neucom.2022.06.081 enhanced with particle swarm optimization technique,” Multimed Tools [21] J. Zheng, Q. Li, and J. Liao, “Heterogeneous type-specific entity Appl, vol. 75, no. 15, pp. 9225–9239, Aug. 2016. representation learning for recommendations in e-commerce network,” https://doi.org/10.1007/s11042-016-3481-4 Information Processing & Management, vol. 58, no. 5, Sep. 2021, Art. [39] W. J. Saunders, S. R. Mousa, and J. Codjoe, “Market basket analysis of no. 102629. https://doi.org/10.1016/j.ipm.2021.102629 safety at active highway-railroad grade crossings,” Journal of Safety [22] F. T. Nobibon, R. Leus, and F. C. R. Spieksma, “Optimization models for Research, vol. 71, pp. 125–137, Dec. 2019. targeted offers in direct marketing: Exact and heuristic algorithms,” https://doi.org/10.1016/j.jsr.2019.09.002 European Journal of Operational Research, vol. 210, no. 3, pp. 670–683, [40] K. Sherly and R. Nedunchezhian, “A improved incremental and May 2011. https://doi.org/10.1016/j.ejor.2010.10.019 interactive frequent pattern mining techniques for market basket analysis [23] I. Viktoratos, A. Tsadiras, and N. Bassiliades, “A context-aware web- and fraud detection in distributed and parallel systems,” Indian Journal of mapping system for group-targeted offers using semantic technologies,” Science and Technology, vol. 8, no. 18, pp. 1–12, 2015. Expert Systems with Applications, vol. 42, no. 9, pp. 4443–4459, Jun. https://doi.org/10.17485/IJST/2015/V8I18/55109 2015. https://doi.org/10.1016/j.eswa.2015.01.039 [41] A. Valarmathi, “Market basket analysis for mobile showroom,” [24] K. Sajithra and R. Patil, “Social media – History and components,” IOSR- International Journal for Research in Applied Science & Engineering JBM, vol. 7, no. 1, pp. 69–74, Jan.–Feb. 2013. Technology, vol. 5, no. X, Oct. 2017. https://doi.org/10.9790/487X-0716974 https://doi.org/10.22214/IJRASET.2017.10185 [25] D. Darshana and C. Vyas, “Problem of customer information overload [42] M. A. Valle, G. A. Ruz, and R. Morrás, “Market basket analysis: and interactive marketing as its solution,” Synergy Journal of Complementing association rules with minimum spanning trees,” Expert Management, vol. 1, pp. 10–16, Jan. 2011. Systems with Applications, vol. 97, pp. 146–162, May 2018. [26] P. K. Singh, P. K. D. Pramanik, and P. Choudhury, “Collaborative https://doi.org/10.1016/j.eswa.2017.12.028 filtering in recommender systems: Technicalities, challenges, [43] H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation st applications, and research trends,” in New Age Analytics, 1 ed., systems: Recommendation models, techniques, and application fields,” G. Shrivastava, S.-L. Peng, H. Bansal, K. Sharma, and M. Sharma, Eds. Electronics, vol. 11, no. 1, Jan. 2022, Art. no. 141. Apple Academic Press, 2020. https://doi.org/10.1201/9781003007210-8 https://doi.org/10.3390/electronics11010141 [27] J. Cao, Z. Wu, Y. Wang, and Y. Zhuang, “Hybrid collaborative filtering [44] H. Shorman and Y. Jbara, “An improved association rule mining algorithm for bidirectional Web service recommendation,” Knowl. Inf. algorithm based on Apriori and Ant Colony approaches,” IOSR Journal Syst., vol. 36, no. 3, pp. 607–627, Sep. 2013. of Engineering, vol. 07, pp. 18–23, Jul. 2017. https://doi.org/10.1007/s10115-012-0562-1 https://doi.org/10.9790/3021-0707011823 [28] P. B.Thorat, R. Goudar, and S. Barve, “Survey on collaborative filtering, [45] N. Zeng and H. Xiao, “Inferring implications in semantic maps via the content-based filtering and hybrid recommendation system,” Apriori algorithm,” Lingua, vol. 239, Feb. 2020, Art. no. 102808. International Journal of Computer Applications, vol. 110, no. 4, pp. 31– https://doi.org/10.1016/j.lingua.2020.102808 36, Jan. 2015. https://doi.org/10.5120/19308-0760 [46] R. Agrawal and R. Srikant, “Fast algorithms for mining association rules [29] F. Bo and C. Jiujun, “Collaborative filtering and recommendation in large databases,” in Proceedings of the 20th International Conference algorithm based on multiple similarities among users,” J. Computer on Very Large Data Bases, San Francisco, CA, USA, Sep. 1994, pp. 487– Science, vol. 39, pp. 23–26, 2012. 499. [47] J. Deshmukh and U. Bhosle, “Image mining using association rule for medical image dataset,” Procedia Computer Science, vol. 85, pp. 117– 124, Jan. 2016. https://doi.org/10.1016/j.procs.2016.05.196 Applied Computer Systems _________________________________________________________________________________________________2022/27 [48] E. M. H. Saeed and B. A. Hammood, “Estimation and evaluation of [62] J. Romero, R. van der Lans, and B. Wierenga, “A partially hidden Markov students’ behaviors in E-learning environment using adaptive model of customer dynamics for CLV measurement,” Journal of computing,” Materials Today: Proceedings, May 2021. Interactive Marketing, vol. 27, no. 3, pp. 185–208, Aug. 2013. https://doi.org/10.1016/j.matpr.2021.04.519 https://doi.org/10.1016/j.intmar.2013.04.003 [49] M. John and H. Shaiba, “Apriori-based algorithm for Dubai road accident [63] J. Benesty, J. Chen, Y. Huang, and I. Cohen, “Pearson correlation analysis,” Procedia Computer Science, vol. 163, pp. 218–227, Jan. 2019. coefficient,” in Noise Reduction in Speech Processing. Springer Topics in https://doi.org/10.1016/j.procs.2019.12.103 Signal Processing, vol 2. Springer, Berlin, Heidelberg, 2009, pp. 1–4. [50] U. Çiçekli and İ. Kabasakal, “Market basket analysis of basket data with https://doi.org/10.1007/978-3-642-00296-0_5 demographics: A case study in E-retailing,” Alphanumeric Journal, [64] K. H. Kim, E. Ko, B. Xu, and Y. Han, “Increasing customer equity of vol. 9, no. 1, pp. 1–12, Jun. 2021. luxury fashion brands through nurturing consumer attitude,” Journal of https://doi.org/10.17093/alphanumeric.752505 Business Research, vol. 65, no. 10, pp. 1495–1499, Oct. 2012. [51] P. Harrington, Machine Learning in Action. Manning, 2012. https://doi.org/10.1016/j.jbusres.2011.10.016 [52] H.-K. Lin, C.-H. Hsieh, N. Wei, and Y. Peng, “Association rules mining [65] F. F. Reichheld and T. Teal, The Loyalty Effect the Hidden Force behind in R for product performance management in industry 4.0,” Procedia Growth, Profits, and Lasting Value. Harvard Business School Press, Jan. CIRP, vol. 83, pp. 699–704, 2019. 1996. https://doi.org/10.1016/J.PROCIR.2019.04.099 [66] M. Utz, S. Johanning, T. Roth, T. Bruckner, and J. Strüker, “From [53] R. Agrawal and R. Srikant, “Fast algorithms for mining association rules,” ambivalence to trust: Using blockchain in customer loyalty programs,” in Readings in Database Systems, 3rd ed. San Francisco, CA, USA: International Journal of Information Management, vol. 68, Mar. 2022, Morgan Kaufmann Publishers Inc., 1998, pp. 580–592. Art. no. 102496. https://doi.org/10.1016/j.ijinfomgt.2022.102496 [54] M. A. Mahdi, K. M. Hosny, and I. Elhenawy, “FR-Tree: A novel rare [67] D. Di Fatta, D. Patton, and G. Viglia, “The determinants of conversion association rule for big data problem,” Expert Systems with Applications, rates in SME e-commerce websites,” Journal of Retailing and Consumer vol. 187, Jan. 2022, Art. no. 115898. Services, vol. 41, pp. 161–168, Mar. 2018. https://doi.org/10.1016/j.eswa.2021.115898 https://doi.org/10.1016/j.jretconser.2017.12.008 [55] T. Kim, D. Kim, and Y. Ahn, “Instant customer base analysis in the financial services sector,” Expert Systems with Applications, vol. 202, Doae Mensouri received the Diploma degree in Computer System and Apr. 2022, Art. no. 117326. https://doi.org/10.1016/j.eswa.2022.117326 Decision Support from the University of Science and Technology UAE [56] K. Jerath, P. S. Fader, and B. G. S. Hardie, “New perspectives on customer Morocco, in 2018. She is in her final year as a PhD student at the UAE ‘death’ using a generalization of the Pareto/NBD model,” 2011. [Online]. University, Tangier, Morocco. She has a publication in the International Journal Available: of Advanced Computer Science and Applications 2022, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=995558 HTTPS://DOI.ORG/10.14569/ijacsa.2022.0130658. [57] X. Chen, R. Geng, and S. Cai, “Predicting microblog users’ lifetime Her research interests include IT, marketing, business intelligence, artificial activities – A user-based analysis,” Electronic Commerce Research and intelligence, decision making, and customer relations. Applications, vol. 14, no. 3, pp. 150–168, May 2015. E-mail: doae.mensouri@etu.uae.ac.ma https://doi.org/10.1016/j.elerap.2014.06.001 ORCID iD: https://orcid.org/0000-0003-2082-2549 [58] J. Dipak and S. S. Siddhartha, “Customer lifetime value research in marketing: A review and future directions,” 2002. [Online]. Available: Abdellah Azmani received his PhD degree in Industrial Computing in https://www.proquest.com/openview/1f7fc325cb76ec2ede9989ff9defb1 Dynamic System Modelling and Artificial Intelligence at the University of 77/1?pq-origsite=gscholar&cbl=32002. Accessed on: Dec. 06, 2021. Science and Technology of Lille in 1991. He worked as a Professor at the Ecole [59] T. Reutterer, M. Platzer, and N. Schröder, “Leveraging purchase Centrale de Lille, France and at the Institute of Computer and Industrial regularity for predicting customer behavior the easy way,” International Engineering from Lens, France. He is a Professor at the Faculty of Science and Journal of Research in Marketing, vol. 38, no. 1, pp. 194–215, Mar. 2021. Technology of Tangier, Morocco. He is a member of the Laboratory of https://doi.org/10.1016/j.ijresmar.2020.09.002 Informatics, System and Telecommunication (LIST) and he has established the [60] R. Colombo and W. Jiang, “A stochastic RFM model,” Journal of Intelligent Automation team, which he coordinates. He has contributed to many Interactive Marketing, vol. 13, no. 3, pp. 2–12, Aug. 1999. theses, scientific research projects, as well as he elaborates and produces many https://doi.org/10.1002/(SICI)1520-6653(199922)13:3<2::AID- IT and decision support solutions for public administration, business DIR1>3.0.CO;2-H management, marketing and logistics. [61] P. S. Fader and B. G. S. Hardie, “The Gamma-Gamma model of monetary E-mail: a.azmani@uae.ac.m value,” 2013. [Online]. Available: http://brucehardie.com/notes/025/ http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Applied Computer Systems de Gruyter

A New Marketing Recommendation System Using a Hybrid Approach to Generate Smart Offers

Applied Computer Systems , Volume 27 (2): 10 – Dec 1, 2022

Loading next page...
 
/lp/de-gruyter/a-new-marketing-recommendation-system-using-a-hybrid-approach-to-19QDkOx00W

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher
de Gruyter
Copyright
© 2022 Doae Mensouri et al., published by Sciendo
ISSN
2255-8691
eISSN
2255-8691
DOI
10.2478/acss-2022-0016
Publisher site
See Article on Publisher Site

Abstract

Applied Computer Systems ISSN 2255-8691 (online) ISSN 2255-8683 (print) December 2022, vol. 27, no. 2, pp. 149–158 https://doi.org/10.2478/acss-2022-0016 https://content.sciendo.com A New Marketing Recommendation System Using a Hybrid Approach to Generate Smart Offers 1* 2 Doae Mensouri , Abdellah Azmani 1,2 Intelligent Automation Laboratory, FST of Tangier Abdelmalek Essaadi University, Tetouan, Morocco Abstract – In order to increase sales, companies try their best to personalized global offers is one of the most efficient ways develop relevant offers that anticipate customer needs. One way to used. A global offer in marketing is defined as the combination achieve this is by leveraging artificial intelligence algorithms that of the main offer with a secondary offer. The main offer process data collected based on customer transactions, extract corresponds to a product or service that meets a basic customer insights and patterns from them, and then present them in a user- need. Whereas the secondary offering, also called an associated friendly way to human or artificial intelligence decision makers. offer, is the equivalent of the service or product offered along This study is based on a hybrid approach, it starts with an online marketplace dataset that contains many customers’ purchases and with the main one. Who is the customer? What will they buy? ends up with global personalized offers based on three different When will they buy it? And how much will they spend? These datasets. The first one, generated by a recommendation system, are classic questions that companies should answer to face the identifies for each customer a list of products they are most likely competition. Thus, the objective of the hybrid approach to buy. The second is generated with an Apriori algorithm. Apriori proposed in this study, which is an extension of a previous work is used as an associate rule mining technique to identify and map [2], is to generate global offers by identifying, for each frequent patterns based on support, confidence, and lift factors, and also to pull important rules between products. The third and customer, a list of the most likely to purchase products, the last one describes, for each customer, their purchase probability probability of purchase in the next few weeks, and the average in the next few weeks, based on the BG/NBD model and the transaction value. Furthermore, to make the most complete average of transactions using the Gamma-Gamma model, as well offers on the market, the associated products that are generally as the satisfaction based on the CLV and RFMTS models. By bought with the initial ones are determined as illustrated in combining all three datasets, specific and targeted promotion Fig. 1. strategies can be developed. Thus, the company is able to anticipate customer needs and generate the most appropriate offers for them while respecting their budget, with minimum II. CONTEXT operational costs and a high probability of purchase The success of any business mainly depends on an attractive transformation. marketing offer and an appropriate call to action. To facilitate Keywords – Apriori algorithm, BG/NBD Gamma-Gamma CLV the task of the customers, to attract more prospects and increase & RFMTS models, marketing recommendation system, smart company’s sales, it is very important to have a good business offer generation. strategy. In this section, the context and the utility of this work, which generate relevant and impactful marketing offers to I. INTRODUCTION attract the attention of target customers, are presented. Selling Nowadays, consumers are particularly concerned about the cannot be improvised; a client cannot be forced to buy. To sell, quality of the services and products offered. They are more customers’ needs should be first identified in order to offer a loyal to brands that make them feel pampered and privileged. product or service that meets their needs. To differentiate from However, the loyalty of customers is critical to the prosperity competitors, a global marketing offer is the key to success. A and success of a business [1]. Thus, companies are increasingly global offer combines the main offer that solves customer needs trying to anticipate customer needs and generate personalized with a secondary offer that contains associated items. product or service offers, which makes it possible to initiate, In this modern world, massive data can lead to information convert and build a relationship with customers. In order to overload problems for digital customers [3]. The consumer keep customers and increase their satisfaction, a marketing feels unable to find the suitable product among the multitude of strategy must extend far beyond conventional usage patterns, items offered on the online marketplace. To overcome this such as customer segmentation, addressing customers by first problem and help the consumer in decision-making, name in a promotional email, or showing relevant recommendation systems (RSs), proven to be a useful advertisements on a website. To provide a truly personalized technology [4], have been developed. These systems reduce the customer experience, companies use more innovative and number of options that the customer has to choose from by intelligent ways to customize the online experience; generating analysing the needs or behaviours. In general, recommendation Corresponding author’s e-mail: doae.mensouri@etu.uae.ac.ma ©2022 Doae Mensouri, Abdellah Azmani. This is an open access article licensed under the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0). Applied Computer Systems _________________________________________________________________________________________________2022/27 Fig. 1. Process of the offer generation approach. systems aim at helping customers make the right purchase presented in Section IV. Finally, Section V discusses the decisions [5]–[7] by analysing and aggregating suggestions conclusions and future research areas. from other customers [8]–[10]. Many websites, such as III. IMPLEMENTATION AliExpress, Netflix, YouTube, and Spotify, use recommendation-based algorithms [11]. The wide application Instead of identifying targeted customers for a specific offer, of those recommender systems, based on either the deep this study takes the opposite route and generates offers by learning approach or the machine learning approach, in the e- anticipating customer needs. As an effective tool to help commerce marketplace is of increasing interest to companies develop targeted marketing strategies, the approach researchers [12]–[21]. While some studies have focused on presented in this paper, as shown in Fig. 1, exploits artificial targeted offers, i.e., the selection of customers who will receive intelligence techniques deployed in parallel sections to create an offer consisting of one or more products during the campaign personalized offers by: [22], [23], this study generates personalized offers that best • Building a recommendation system by anticipating match the preferences of each customer. customer needs based on historical data, then generating Consumers are constantly bombarded, mainly by email [24], association rules to identify frequent patterns between with proposals for offers that often do not correspond to their items and to identify associated ones; profile or their needs. On the supplier’s side, this translates into • Identifying the probability of purchase in the next 4, 8, and additional costs since it is necessary to send several thousand 12 weeks; emails to be able to achieve poor results and into a questioning • Predicting the value of next purchase; of its image because a large number of people perceive its • Calculating customer’s profitability and satisfaction; proposals as an annoyance and a saturation of their email [25]. • Generating a global, personalized offer. It is to remedy such problems that the approach proposed in the A. About the Used Dataset context of this article is inscribed and which generates an irresistible and personalized offer that best matches the The dataset used in this article is called “Brazilian E- customer’s preferences, thus, on the one hand, providing value Commerce Public Dataset”, provided by Olist Store and and enticing customers to purchase and, on the other hand, accessible on the Kaggle website (https://www.kaggle.com). efficiently reducing the problem of information overload. The This dataset contains multiple purchases from 2016 to 2018 at delivered offers have a high transformation factor due to the various stores using the Olist service. In this study, only the combination of several methods such as recommendation Customers (99 441 rows), Orders (99 478 rows), Reviews systems, the Apriori algorithm, association rules, the BG/NBD (98 410 rows) and Payments (99 478 rows) sub-datasets are model, and the Gamma-Gamma model, as explained in the used since they contain the useful variables for this study following sections. approach. The three sub-datasets are filtered and prepared, then This article is organised as follows: Section II provides an merged by their respective primary keys into a single initial overview of the context related to marketing offer generation. dataset as shown in Table I. Section III presents the construction and evaluation of the In order to identify the probability of future purchases in the various models used. A discussion of the obtained results is next few weeks, the value of the next one, as well as customer Applied Computer Systems _________________________________________________________________________________________________2022/27 satisfaction, five variables, which are recency (R), frequency neighbours to recommend an article to a target customer [3], (F), monetary value (M), interpurchase time (T), and [27]. satisfaction (S), should be first calculated. To do so, the RFMTS Collaborative Filtering Model (CF) model [2] is used, Table II presents the dataset after generating Recommendation systems often use collaborative filtering. It the R, F, M, T, and S variables for each customer. builds on the idea that the best recommendations come from • R: represents the period between the current moment and other customers with similar taste. This means using similar the customer’s last purchase. A low recency rate indicates customer review history to predict how a customer will rate a that customers buy frequently. product. CF models can recommend a product to any user X • F: represents the number of transactions made by a based on the preferences of a like-minded user Y [28]. customer. A high frequency rate indicates a high level of Collaborative filtering has two subcategories: model-based and loyalty. memory-based. This study covers collective memory-based • M: represents the average amount of a specific customer’s filtering given that user-based (user-user) and item-based (item- purchases. A high monetary value implies that the to-item) filtering belong to this category. The difference customer brings high profits to the company. between the two mainly lies in the considerations when • T: represents the time between a customer’s two calculating the recommendation. successive purchases. • S: represents the total of a customer’s ratings divided by • User-based collaborative filtering: it focuses on the the number of reviews. “nearest neighbour” approach [29] to recommend items by looking at user ratings to find the “nearest neighbours” TABLE I with ratings similar to the targeted customer. In other INITIAL DATASET words, user-based filtering finds users with similar Order Customer Payment Review spending habits and provides recommendations based on Order ID purchase ID value score time their interests. 10/2/2017 4 • Item-based collaborative filtering: first introduced by e4…f7 10.11 000…1e2 10:56 Amazon in 1998 [30]. Contrary to user-based 10/2/2017 4 collaborative filters, item-based filtering finds patterns of 9c…3c 18.12 000…be3 10:56 similarity between different elements by considering the 10/2/2017 3 number of users who have purchased items A and B 11…62 25.59 000…064 10:56 simultaneously. If the correlation between item A and B is … … … … high enough, it can be assumed that the two items are 142.14 4 7/24/2018 similar to each other. As a result, customers who purchase 53…51 fff…684 20:41 item B will receive a recommendation of item A and vice 8/8/2018 1 versa. 47…5d 20.46 fff…30a 8:38 Fig. 2 shows the various collaborative filtering approaches TABLE II based on the user and the item. The dotted arrows on the left DATASET AFTER CALCULATING R, F, M, T AND S VALUES represent recommendations based on user similarity and Customer preference, and the dotted arrows on the right represent R F M T S ID recommendations based on similar items. 000…1e2 1428.77 2.10 120.66 432.67 1.20 Evaluation 000…be3 230.48 2.24 180.04 345.89 3.15 In machine learning, the accuracy of a model is evaluated by 000…064 428.96 1.63 80.23 115.66 4.22 mathematically taking the root mean square of the error that has … … … … … … occurred between the test and the predicted values [31]. This fff…684 149.52 1.96 71.99 109.71 4.64 error value is used as a defensive score to decrease the fff…30a 230.41 2.15 44.84 84.32 3.27 individual effects of individual measurement errors so that there are no significant assumptions that substantially distort the B. Recommendation Systems results [11]. The RMSE (Root Mean Square Error) is a method In recommender systems, two algorithms are mainly used: for determining the accuracy of a predictive target value model content-based filtering (CBF) and collaborative filtering (CF). [32]. It is the estimate of the average difference between the The CBF technique is primarily based on customer and product actual value and the estimated value, which is calculated as information. It recommends a product to a customer based on shown in (1). Where n is the number of estimates, yj represents the profile, feedback, and item information, etc. Its main the current estimate, and ŷj represents the average value of the restriction is the limited analysis of content as well as the estimates. The comparison of the user-based model and item- specialization [26]. CF recommends products to consumers based model results based on the RMSE value is represented based on their previous ratings to alleviate the above problems. Table III. This shows that the user-based model gave a better This technique is based on the preferences of the top k similar result, with a minimum RMSE error value. Applied Computer Systems _________________________________________________________________________________________________2022/27 Fig. 2. User-based and item-based collaborative filtering approaches. [45]. Rules, whose lift is greater than 1, are the practicable association rules, which implies that item A has a positive effect 1 2 RMSE YY− . (1) on the appearance of item B. The support for a set of items is ( ) ∑ jj j=1 the percentage of the dataset that contains this set of items [50], [51]. Let I = {i , i ,...,i } be a set of items and A – a set of some a b z The implementation of the user-based model on the dataset items in I. Equation (2) defines the support of A, where identifies the top 5 recommended items for each customer, Count (A) is the number of times item-set A appears in all based on customer’s preferences (as shown in Table IV). transactions and N is the number of transactions. TABLE III C. Generating Association Rules by Applying Apriori Algorithm USER-BASED AND THE ITEM-BASED MODELS RESULTS BASED ON THE RMSE Market basket analysis (MBA) is a data mining technique VALUE commonly used to identify which products tend to be purchased Model RMSE together [33], [34] and predict customer purchasing behaviour. user-based model 1.063122817661801 It discovers the association between the activities performed by item-based model 1.343641161111319 the customers [35]. Although the primary focus of MBA is on shopping carts and customers [36]–[38], it also includes fields such as highway-railroad safety [39], fraud detection [40], Count (A) mobile showroom [41], and others. MBA tries to identify Support (A) = . (2) 𝑁𝑁 purchase patterns in a large number of customer transactions. This is where association rules play an essential role by finding For rule A⇒B (antecedent →consequent), the Support, relations between different items and discovering frequent ones Confidence and Lift equations are the same presented in [52]: [34], [42]. Association rules do not capture individual preferences but instead look for relationships between different • The support of the A→B association rule is the sets of elements in each transaction. This is what makes them percentage of transactions that include both elements different from collaborative filtering, which is used in A and B. It is defined as shown in (3). recommendation systems [43]. To obtain the networks of {A,B} products, the Apriori algorithm, the most frequent and efficient Support (A → B) = . (3) 𝑁𝑁 algorithm for association rule generation [44], [45], is applied in this research to generate these rules. • Confidence measures the number of times items in B The Apriori algorithm was the first to determine Boolean appear in transactions involving A [53]. The association rules by extracting patterns from datasets [46]. It is Confidence is calculated as shown in (4). used in different fields, in medicine to discover frequent image patterns in mammogram images [47], in education to predict {A,B} Confidence (𝐴𝐴 → B) = . (4) Support (A) system of learning [48], in traffic accidents to identify the main causes and trends related with it [49]. The Apriori algorithm is • Lift is a probability that event B occurs independently a method to find recurring items in a transactional dataset and as a consequence of another event A [35]. to generate associated rules. It first determines frequent, unique Equation (5) defines the lift of the association rule items through transactions, then generates association rules A→B. [45]. The set of frequently occurring items corresponds to the set of frequently occurring functions measured by the parameter {A,B} Lift (A → B) = . (5) support. Confidence and lift parameters are used to measure the Support (B) association rules corresponding to implication rules (A→B) = Applied Computer Systems _________________________________________________________________________________________________2022/27 The traditional approach to association rules is to mine each customer transaction set produced meaningful association records for common item-sets with a minimum threshold of rules with high confidence (greater than 0.7) as shown in “support” and then use another minimum threshold of Table V. These results, which determine which products are “confidence” to generate rules for the collected common item- frequently bought together, allow companies to develop sets [54] (Fig. 3). Applying the priori association algorithm to customer loyalty strategies by generating global offers. Fig. 3. Association rule mining. The application of the BG/NBD model to this article dataset, TABLE IV including the RFT data generated previously, makes it possible TOP 5 RECOMMENDATIONS FOR EACH CUSTOMER to create the Frequency-Recency Matrix shown in Fig. 4. The Customer ID Top 5 recommended items visualization of this matrix, over four weeks (Fig. 4), shows the B00CGW74YU (Laptop), probability of future customer purchases at a given time. This B000Q8UAWY (Phone), probability increases if the number of purchases and the 000…1e2 B003TXZCYE (Perfume), B008DW95NA (T-shirt), duration since its last purchase are important. It should be noted B005N3TPJQ (Gaming mouse) that the probabilities of future purchases are quite low over the B000P6G7YW (Perfume), next four weeks and are limited to a few customers; this is B00000JX3C (Baby toy), explained by the pattern of sporadic purchases seen in the 000…be3 B000GC2DDE (Baby clothes), studied dataset. B001PB17YY (Backpack), B0022TQXWG (Diaper) Based on the previous model, predictions could be made for any chosen period. In the case of this study, predictions are B0002KR13M (Dress), B000VAIMF4 (Sunscreen), created for the first 4, 8, and 12 weeks, as these are a strong 000…064 B002ZYQE0Y (Perfume), short-term indicator for focusing a vendor’s efforts on their B0058SYDTI (Swimsuit), customers with the highest likelihood of future purchases. B004TLL730 (Makeup) Table VI shows the probability of purchase in the next 4, 8, and … … 12 weeks for each customer. TABLE V BG/NBD Validation ASSOCIATION RULES Model predictions cannot be used blindly, they must be Rule Antecedents Consequents Support Conf Lift validated with reality. For this purpose, the last three months of 0 e0c…a26 368…2db 0.37 0.79 1.68 results, when the numbers of transactions of each customer are … … … … … … already known, are compared with the BG/NBD model results. 125 389…0f4 0bc…cdd 0.40 0.84 1.68 As can be seen in Fig. 5, the lines are quite close to each 126 0bc…894 537…48c 0.39 0.80 1.70 other, which means that the model makes predictions quite 127 537…f73 0bc…48c 0.41 0.86 1.70 close to the margin of error. In all cases, the model is close on certain points to reality. Similarly, extending the duration of the … … … … … … test could improve the validation of the model. E. Predictive Calculation of the Future Purchase Value by D. Calculation of the Probability of a Future Purchase by Application of the Gamma-Gamma Model Application of the BG / NBD Model The probability of having each customer’s purchases forecast The BG/NBD model [55] is used in this study to estimate, for for the following months can be known, but not their number. each customer, the probability of carrying out transactions in a A customer who buys 20 times on average 5 USD is not the future period. The BG/NBD is a simplified substitutional model same as another who buys 2 times 400 USD. It is at this stage to the well-known Pareto/NBD model [56], which is based on that the Gamma-Gamma model initially proposed by Colombo Bayesian probability in a hierarchical way to make its and Jiang [60] is used to give an estimate of the profitability of estimates. These two models are particularly used in marketing each customer, using the monetary value (M) column research, especially in the analysis of the customer database. previously generated. They are rooted in the theory of customer lifetime value [57]– One of the requirements of the Gamma-Gamma model is to [59] that forecasts future customer behaviour based on only work with observations with a purchase frequency stochastic aspects of interaction behaviour, such as different from 0 (retain only customers with more than one departure [56]. purchase) [61]. Similarly, it must be ensured that there is no Applied Computer Systems _________________________________________________________________________________________________2022/27 TABLE VI correlation between the frequency (F) and the monetary PURCHASE PREDICTIONS OF NEXT 4, 8, AND 12 WEEKS value (M) of each consumer [62]. In practice, the Gamma- Gamma model must check whether the Pearson correlation [63] Expected Expected Expected Customer ID 4 weeks 8 weeks 12 weeks between the two vectors is close to 0 (Fig. 6). 000…1e2 0.02 0.04 0.05 One of the faculties of this model is to calculate the conditional expectation of the average profit per transaction for 000…be3 0.03 0.07 0.01 each customer as shown in Table VII. 000…064 0.02 0.04 0.07 … … … … F. Customer Profitability and Satisfaction fff…84 0.02 0.04 0.07 Customer Lifetime Value (CLV or CLTV) is a marketing metric [64] that, in a nutshell, reflects the profit that a certain ffff371b4d645b6ecea244b27531430a 0.02 0.01 0.01 customer will provide to a business over a given period of time. TABLE VII This makes it possible to get out of the subjectivism of the value of a customer, in order to quantify it in a more exploitable field CUSTOMER’S NEXT AVERAGE TRANSACTION VALUE linked to the profitability it brings to the company. The concept Customer ID Avg transaction of customer lifetime value is well accepted by researchers and 0000366f3b9a7992bf8c76cfdf3221e2 12 398.64 marketers [58] who agree that long-term customers are more 0000b849f77a49e4a4ce2b2a4ca5be3 4542.78 profitable. The reason commonly given for increasing the 0000f46a3911fa3c0805444483337064 6608.59 profitability of long-term customers is the increase in profit … … from loyal customers, which is determined by the price fffcf5a5ff07b0908bd4e2dbc735a684 344.3 premium they pay, the additional profit from sales through referrals, and the benefits in relation to cost savings due to the ffff371b4d645b6ecea244b27531430a 69.25 increased sales in their regard [65]. TABLE VIII NORMALIZED CLV AND S Customer unique id CLV S 0000366f3b9a7992bf8c76cfdf3221e2 0.928328 0.05 0000b849f77a49e4a4ce2b2a4ca5be3 0.870307 0.5375 0000f46a3911fa3c0805444483337064 0.662116 0.805 … … fffcf5a5ff07b0908bd4e2dbc735a684 0.098976 0.91 ffff371b4d645b6ecea244b27531430a 0.095563 0.5675 Customer satisfaction and trust are respectively the first and second important antecedents of customer loyalty [66]. After calculating the CLV for each customer, based on the RFMTS model [2], the CLV and satisfaction (S) values are normalized to a range between 0 and 1, using the well-known statistic normalization formula (6), as indicated in Table VIII. 𝑋𝑋𝑋𝑋 −min(𝑋𝑋 ) Fig. 4. Frequency recency matrix. 𝑋𝑋 norm = . (6) ( ) max 𝑋𝑋 −min(𝑋𝑋 ) Fig. 6. Frequency & Monterey correlation matrix. Fig. 5. Actual purchases in holdout period vs. predicted purchase. Applied Computer Systems _________________________________________________________________________________________________2022/27 IV. RESULTS AND DISCUSSION: HOW TO CREATE A GLOBAL algorithm, so generation is automatic which takes as input OFFER BASED ON DIFFERENT INFORMATION GENERATED information about customers and their purchases and outputs PREVIOUSLY generated offers. The global algorithm developed in this study makes it possible to implement parallel sections of the process Predicting future customer behaviour and generating a simultaneously to reduce processing time, and then the results dedicated global offer for each of them allows the company to of each section are combined to provide a personalized offer. consolidate its customer relationships, differentiate itself from its competitors, and increase its earnings. It is in this context V. VALIDATION that the study presented here takes place, with the aim of anticipating customer needs in order to generate a global offer In addition to the different evaluation factors concerning by following the steps illustrated in Fig. 1. The global offer is each technique used at different stages of this study, the offer made up of the main offer and associated offer for customers transformation rate is used in the end to validate and evaluate who are probably likely to accept the offer positively based on the proposed offer generation algorithm. The conversion rate is their needs. The main offer corresponds to a product or service a key factor in the financial success of a platform [67]. It is a that satisfies consumer’s basic needs. While associated offer is percentage metric that helps companies analyse the added to the main one with additional products or services that performance of their marketing strategies and can be calculated are generally purchased with the bought one. The global offer by dividing the number of sales by the total number of creation is divided into two steps. The first step is the main offer communicated offers and multiplying by 100 %, as shown creation; in this step a recommendation system is used to in (7). identify, for each customer, the top 5 products that meet the Number of sales customer needs (Table IV). The second step is the associated Conversion rate= · 100. (7) Number of communicated offers offer creation; in this step Apriori algorithm is used to determine associated products as shown in Table V. After To validate our approach, the data are split into two parts, determining the products of interest to a specific customer, as with the last 20 % used for the validation and 80 % used for well as the associated products, the third step takes place, which testing. The analysis of the 80 % of data has made it possible to aims to calculate the probability for a customer to transact in a determine for each customer the products of interest, the future period using the BG/NBD model. The probabilities of associated products, purchase time and purchase amount, and customer’s future purchases in the next 4, 8 and 12 weeks give the value of CLV and satisfaction. Let us take as an example a an idea of the ideal time to communicate to the customer a set of 65 customers identified by our model as being interested personalized offer. With the above, the probability of each in the product “laptop” and that their CLV and satisfaction customer purchase in the next few months is known but not its scores are close. Fig. 1 represents the personalized offer amount. This is where the role of the next step comes in, which generated for those 65 customers. The conversion rate value of consists in calculating the future purchase value by application this offer, calculated by (7), is 70 %, which is a high value that of the Gamma-Gamma model. Knowing the value of the next means an average of 45 customers out of 65 positively purchase helps choose the associated products, identified in interacted with the generated offer. Step 2, while respecting the customer’s budget, in other words, the top products, the total price of which does not exceed the VI. CONCLUSION client’s next purchase value, are chosen. Combining all the Recent advances in technology are impacting not just information generated in the above steps makes it possible to consumers, but merchants as well. On the one hand, today’s generate a global personalized offer with a high transform consumers are constantly changing, and digitization is the value. All customers are not the same, and so they should not major reason for this change in behaviour. On the other hand, be treated the same way. A loyal customer deserves eventually AI algorithms have given marketers the ability to process and more promotions and services added to the offer. Thus, to extract insights from collected customer transaction data and identify the value of a customer, it is necessary to calculate the thereby increase profitability with minimum operational cost, CLV value (calculated based on customer recency, frequency which is the goal of any type of business. Hence, this article has and monetary value) and satisfaction value (calculated based on represented a process of generating personalized offers as part customer review rating). By following the approach proposed of an overall marketing strategy aimed at anticipating customer in this study, a global offer that contains associated products needs and developing targeted and specific promotional and personalized services is generated for each customer. In strategies. First, a recommendation system is built based on addition to the probability of purchase, the global offer respects customer historical data, and then association rules are the customer’s budget as well. Fig. 7 represents an example of generated to identify frequent patterns between items. After a personalized offer generated based on the approach described determining what each customer may buy, the next step is to above for a specific customer, where “Laptop” is the main determine the best time and the value of the next purchase. customer’s need, and the remaining items are additional ones Finally, the profitability and satisfaction of each customer are with the additional option of checking and unchecking the items identified. proposed while adapting the price to make it attractive. All Due to the approach presented in this study, customer methods used in this study are integrated into a global overload problems are minimised since the customers receive Applied Computer Systems _________________________________________________________________________________________________2022/27 personalized offers about products that meet their needs. approach presented in this paper can be reinforced by taking Competition is getting tougher, and to differentiate themselves into consideration other parameters, in particular the probability from competitors and increase the conversion rate of offers, of purchase, the profile of the customer, as well as the budget, companies must forget about old marketing strategies and opt to plan the sending of the offer while respecting the customer’s for personalization. In addition to the customization of the schedule and personalizing the associated message according to offers, which is the objective of this study, choosing the right the profile (age, gender, residence, etc.). moment and the right sending channel to send the offer increases customer satisfaction. Thus, in future research, the Fig. 7. Global offer example. [6] N. Jing, T. Jiang, J. Du, and V. Sugumaran, “Personalized CKNOWLEDGMENT recommendation based on customer preference mining and sentiment The research has been supported by the Ministry of Higher assessment from a Chinese e-commerce website,” Electronic Commerce Research, vol. 18, no. 1, pp. 159–179, Nov. 2018. Education, Scientific Research and Innovation, the Digital https://doi.org/10.1007/s10660-017-9275-6 Development Agency (DDA) and the National Centre for [7] H. Zhang, L. Zhao, and S. Gupta, “The role of online product Scientific and Technical Research (CNRST) of Morocco recommendations on customer decision making and loyalty in social shopping communities,” International Journal of Information (Smart DLSP Project – AL KHAWARIZMI IA-PROGRAM). Management, vol. 38, no. 1, pp. 150–166, Feb. 2018. https://doi.org/10.1016/j.ijinfomgt.2017.07.006 REFERENCES [8] E. Frias-Martinez, S. Y. Chen, and X. Liu, “Evaluation of a personalized digital library based on cognitive styles: Adaptivity vs. adaptability,” [1] K. Matsuoka, “Effects of revenue management on perceived value, International Journal of Information Management, vol. 29, no. 1, pp. 48– customer satisfaction, and customer loyalty,” Journal of Business 56, Feb. 2009. https://doi.org/10.1016/j.ijinfomgt.2008.01.012 Research, vol. 148, pp. 131–148, Sep. 2022. [9] E. Frias-Martinez, G. Magoulas, S. Chen, and R. Macredie, “Automated https://doi.org/10.1016/j.jbusres.2022.04.052 user modeling for personalized digital libraries,” International Journal of [2] D. Mensouri, A. Azmani, and M. Azmani, “K-Means customers Information Management, vol. 26, no. 3, pp. 234–248, Jun. 2006. clustering by their RFMT and score satisfaction analysis,” International https://doi.org/10.1016/j.ijinfomgt.2006.02.006 Journal of Advanced Computer Science and Applications (IJACSA), [10] H.-N. Kim, A.-T. Ji, I. Ha, and G.-S. Jo, “Collaborative filtering based on vol. 13, no. 6, 2022, Art no. 6. collaborative tagging for enhancing the quality of recommendation,” https://doi.org/10.14569/IJACSA.2022.0130658 Electron. Commer. Res. Appl., vol.9, no. 1, pp. 73–83, Jan.–Feb. 2010. [3] P. K. Singh, E. Othman, R. Ahmed, A. Mahmood, H. Dhahri, and https://doi.org/10.1016/j.elerap.2009.08.004 P. Choudhury, “Optimized recommendations by user profiling using [11] P. Boström and M. Filipsson, “Comparison of user based and item based apriori algorithm,” Applied Soft Computing, vol. 106, Jul. 2021, Art. no. collaborative filtering recommendation services,” 2017. [Online]. 107272. https://doi.org/10.1016/j.asoc.2021.107272 Available: https://www.semanticscholar.org/paper/Comparison-of-User- [4] B. Tilahun, C. Awono, and B. Batchakui, “A survey of state-of-the-art: Based-and-Item-Based-Filtering-Bostr%C3%B6m- Deep learning methods on recommender system,” IJCA, vol. 162, no. 10, Filipsson/27984027b4fccd7323371129768c459ba16f8fbd. Accessed on: pp. 17–22, Mar. 2017. https://doi.org/10.5120/ijca2017913361 Mar. 04, 2022. [5] I. Y. Choi, M. G. Oh, J. K. Kim, and Y. U. Ryu, “Collaborative filtering [12] Y. Chen, Y. Lu, B. Wang, and Z. Pan, “How do product recommendations with facial expressions for online video recommendation,” International affect impulse buying? An empirical study on WeChat social commerce,” Journal of Information Management, vol. 36, no. 3, pp. 397–402, Jun. Information & Management, vol. 56, no. 2, pp. 236–248, Mar. 2019. 2016. https://doi.org/10.1016/j.ijinfomgt.2016.01.005 https://doi.org/10.1016/j.im.2018.09.002 Applied Computer Systems _________________________________________________________________________________________________2022/27 [13] T. Donkers, B. Loepp, and J. Ziegler, “Sequential user-based recurrent [30] G. D. Linden, J. A. Jacobi, and E. A. Benson, “Collaborative neural network recommendations,” in Proceedings of the Eleventh ACM recommendations using item-to-item similarity mappings,” Conference on Recommender Systems, New York, NY, USA, Aug. 2017, U.S. patent US6266649B1, 1998. [Online]. Available: pp. 152–160. https://doi.org/10.1145/3109859.3109877 https://patents.google.com/patent/US6266649/en. Accessed on: Mar. 04, [14] B. Hidasi and A. Karatzoglou, “Recurrent neural networks with top-k 2022. gains for session-based recommendations,” in Proceedings of the 27th [31] S. Jayalakshmi, N. Ganesh, R. Čep, and J. S. Murugan, “Movie ACM International Conference on Information and Knowledge recommender systems: Concepts, methods, challenges, and future Management, New York, NY, USA, Oct. 2018, pp. 843–852. directions,” Sensors, vol. 22, no. 13, Jan. 2022, Art. no. 13. https://doi.org/10.1145/3269206.3271761 https://doi.org/10.3390/s22134904 [15] I. Islek and S. G. Oguducu, “A hierarchical recommendation system for [32] Ç. Odabaşı, P. Dologlu, F. Gülmez, G. Kuşoğlu, and Ö. Çağlar, E-commerce using online user reviews,” Electronic Commerce Research “Investigation of the factors affecting reverse osmosis membrane and Applications, vol. 52, Mar.–Apr. 2022, Art. no. 101131. performance using machine-learning techniques,” Computers & https://doi.org/10.1016/j.elerap.2022.101131 Chemical Engineering, vol. 159, Mar. 2022, Art no. 107669. [16] Z. Li, H. Zhao, Q. Liu, Z. Huang, T. Mei, and E. Chen, “Learning from https://doi.org/10.1016/j.compchemeng.2022.107669 history and present: Next-item recommendation via discriminatively [33] P. Jirapatsil and N. Phumchusri, “Market basket analysis for fresh exploiting user behaviors,” in Proceedings of the 24th ACM SIGKDD products location improvement: A case study of E-commerce business International Conference on Knowledge Discovery & Data Mining, New warehouse,” in Proceedings of the 4th International Conference on York, NY, USA, Jul. 2018, pp. 1734–1743. Management Science and Industrial Engineering, New York, NY, USA, https://doi.org/10.1145/3219819.3220014 Apr. 2022, pp. 23–28. https://doi.org/10.1145/3535782.3535786 [17] N. Sachdeva and J. McAuley, “How useful are reviews for [34] A. Monteserin and M. G. Armentano, “Influence-based approach to recommendation? A critical review and potential improvements,” in market basket analysis,” Information Systems, vol. 78, pp. 214–224, Nov. Proceedings of the 43rd International ACM SIGIR Conference on 2018. https://doi.org/10.1016/j.is.2018.01.008 Research and Development in Information Retrieval, New York, NY, [35] S. Halim, T. Octavia, and C. Alianto, “Designing facility layout of an USA, Jul. 2020, pp. 1845–1848. amusement arcade using market basket analysis,” Procedia Computer https://doi.org/10.1145/3397271.3401281 Science, vol. 161, pp. 623–629, Jan. 2019. [18] B. Shao, X. Li, and G. Bian, “A survey of research hotspots and frontier https://doi.org/10.1016/j.procs.2019.11.165 trends of recommendation systems from the perspective of knowledge [36] Y. Chen, “Research on personalized recommendation algorithm based on graph,” Expert Systems with Applications, vol. 165, Mar. 2021, Art no. user preference in mobile e-commerce,” Information Systems and e- 113764. https://doi.org/10.1016/j.eswa.2020.113764 Business Management, vol. 18, pp. 837–850, Dec. 2020. [19] D. R. Stöckli and H. Khobzi, “Recommendation systems and convergence https://doi.org/10.1007/s10257-019-00401-2 of online reviews: The type of product network matters!” Decision [37] M. Gupta, S. Kochhar, P. Jain, and P. Nagrath, “Hybrid recommender Support Systems, vol. 142, Mar. 2021, Art. no. 113475. system using A-priori algorithm.” in Proceedings of International https://doi.org/10.1016/j.dss.2020.113475 Conference on Sustainable Computing in Science, Technology and [20] J. Zhang, M. Li, W. Liu, S. Lauria, and X. Liu, “Many-objective Management (SUSCOM), Jaipur, India, Feb. 24, 2019. optimization meets recommendation systems: A food recommendation https://doi.org/10.2139/ssrn.3349290 scenario,” Neurocomputing, vol. 503, pp. 109–117, Sep. 2022. [38] R. Katarya and O. P. Verma, “A collaborative recommender system https://doi.org/10.1016/j.neucom.2022.06.081 enhanced with particle swarm optimization technique,” Multimed Tools [21] J. Zheng, Q. Li, and J. Liao, “Heterogeneous type-specific entity Appl, vol. 75, no. 15, pp. 9225–9239, Aug. 2016. representation learning for recommendations in e-commerce network,” https://doi.org/10.1007/s11042-016-3481-4 Information Processing & Management, vol. 58, no. 5, Sep. 2021, Art. [39] W. J. Saunders, S. R. Mousa, and J. Codjoe, “Market basket analysis of no. 102629. https://doi.org/10.1016/j.ipm.2021.102629 safety at active highway-railroad grade crossings,” Journal of Safety [22] F. T. Nobibon, R. Leus, and F. C. R. Spieksma, “Optimization models for Research, vol. 71, pp. 125–137, Dec. 2019. targeted offers in direct marketing: Exact and heuristic algorithms,” https://doi.org/10.1016/j.jsr.2019.09.002 European Journal of Operational Research, vol. 210, no. 3, pp. 670–683, [40] K. Sherly and R. Nedunchezhian, “A improved incremental and May 2011. https://doi.org/10.1016/j.ejor.2010.10.019 interactive frequent pattern mining techniques for market basket analysis [23] I. Viktoratos, A. Tsadiras, and N. Bassiliades, “A context-aware web- and fraud detection in distributed and parallel systems,” Indian Journal of mapping system for group-targeted offers using semantic technologies,” Science and Technology, vol. 8, no. 18, pp. 1–12, 2015. Expert Systems with Applications, vol. 42, no. 9, pp. 4443–4459, Jun. https://doi.org/10.17485/IJST/2015/V8I18/55109 2015. https://doi.org/10.1016/j.eswa.2015.01.039 [41] A. Valarmathi, “Market basket analysis for mobile showroom,” [24] K. Sajithra and R. Patil, “Social media – History and components,” IOSR- International Journal for Research in Applied Science & Engineering JBM, vol. 7, no. 1, pp. 69–74, Jan.–Feb. 2013. Technology, vol. 5, no. X, Oct. 2017. https://doi.org/10.9790/487X-0716974 https://doi.org/10.22214/IJRASET.2017.10185 [25] D. Darshana and C. Vyas, “Problem of customer information overload [42] M. A. Valle, G. A. Ruz, and R. Morrás, “Market basket analysis: and interactive marketing as its solution,” Synergy Journal of Complementing association rules with minimum spanning trees,” Expert Management, vol. 1, pp. 10–16, Jan. 2011. Systems with Applications, vol. 97, pp. 146–162, May 2018. [26] P. K. Singh, P. K. D. Pramanik, and P. Choudhury, “Collaborative https://doi.org/10.1016/j.eswa.2017.12.028 filtering in recommender systems: Technicalities, challenges, [43] H. Ko, S. Lee, Y. Park, and A. Choi, “A survey of recommendation st applications, and research trends,” in New Age Analytics, 1 ed., systems: Recommendation models, techniques, and application fields,” G. Shrivastava, S.-L. Peng, H. Bansal, K. Sharma, and M. Sharma, Eds. Electronics, vol. 11, no. 1, Jan. 2022, Art. no. 141. Apple Academic Press, 2020. https://doi.org/10.1201/9781003007210-8 https://doi.org/10.3390/electronics11010141 [27] J. Cao, Z. Wu, Y. Wang, and Y. Zhuang, “Hybrid collaborative filtering [44] H. Shorman and Y. Jbara, “An improved association rule mining algorithm for bidirectional Web service recommendation,” Knowl. Inf. algorithm based on Apriori and Ant Colony approaches,” IOSR Journal Syst., vol. 36, no. 3, pp. 607–627, Sep. 2013. of Engineering, vol. 07, pp. 18–23, Jul. 2017. https://doi.org/10.1007/s10115-012-0562-1 https://doi.org/10.9790/3021-0707011823 [28] P. B.Thorat, R. Goudar, and S. Barve, “Survey on collaborative filtering, [45] N. Zeng and H. Xiao, “Inferring implications in semantic maps via the content-based filtering and hybrid recommendation system,” Apriori algorithm,” Lingua, vol. 239, Feb. 2020, Art. no. 102808. International Journal of Computer Applications, vol. 110, no. 4, pp. 31– https://doi.org/10.1016/j.lingua.2020.102808 36, Jan. 2015. https://doi.org/10.5120/19308-0760 [46] R. Agrawal and R. Srikant, “Fast algorithms for mining association rules [29] F. Bo and C. Jiujun, “Collaborative filtering and recommendation in large databases,” in Proceedings of the 20th International Conference algorithm based on multiple similarities among users,” J. Computer on Very Large Data Bases, San Francisco, CA, USA, Sep. 1994, pp. 487– Science, vol. 39, pp. 23–26, 2012. 499. [47] J. Deshmukh and U. Bhosle, “Image mining using association rule for medical image dataset,” Procedia Computer Science, vol. 85, pp. 117– 124, Jan. 2016. https://doi.org/10.1016/j.procs.2016.05.196 Applied Computer Systems _________________________________________________________________________________________________2022/27 [48] E. M. H. Saeed and B. A. Hammood, “Estimation and evaluation of [62] J. Romero, R. van der Lans, and B. Wierenga, “A partially hidden Markov students’ behaviors in E-learning environment using adaptive model of customer dynamics for CLV measurement,” Journal of computing,” Materials Today: Proceedings, May 2021. Interactive Marketing, vol. 27, no. 3, pp. 185–208, Aug. 2013. https://doi.org/10.1016/j.matpr.2021.04.519 https://doi.org/10.1016/j.intmar.2013.04.003 [49] M. John and H. Shaiba, “Apriori-based algorithm for Dubai road accident [63] J. Benesty, J. Chen, Y. Huang, and I. Cohen, “Pearson correlation analysis,” Procedia Computer Science, vol. 163, pp. 218–227, Jan. 2019. coefficient,” in Noise Reduction in Speech Processing. Springer Topics in https://doi.org/10.1016/j.procs.2019.12.103 Signal Processing, vol 2. Springer, Berlin, Heidelberg, 2009, pp. 1–4. [50] U. Çiçekli and İ. Kabasakal, “Market basket analysis of basket data with https://doi.org/10.1007/978-3-642-00296-0_5 demographics: A case study in E-retailing,” Alphanumeric Journal, [64] K. H. Kim, E. Ko, B. Xu, and Y. Han, “Increasing customer equity of vol. 9, no. 1, pp. 1–12, Jun. 2021. luxury fashion brands through nurturing consumer attitude,” Journal of https://doi.org/10.17093/alphanumeric.752505 Business Research, vol. 65, no. 10, pp. 1495–1499, Oct. 2012. [51] P. Harrington, Machine Learning in Action. Manning, 2012. https://doi.org/10.1016/j.jbusres.2011.10.016 [52] H.-K. Lin, C.-H. Hsieh, N. Wei, and Y. Peng, “Association rules mining [65] F. F. Reichheld and T. Teal, The Loyalty Effect the Hidden Force behind in R for product performance management in industry 4.0,” Procedia Growth, Profits, and Lasting Value. Harvard Business School Press, Jan. CIRP, vol. 83, pp. 699–704, 2019. 1996. https://doi.org/10.1016/J.PROCIR.2019.04.099 [66] M. Utz, S. Johanning, T. Roth, T. Bruckner, and J. Strüker, “From [53] R. Agrawal and R. Srikant, “Fast algorithms for mining association rules,” ambivalence to trust: Using blockchain in customer loyalty programs,” in Readings in Database Systems, 3rd ed. San Francisco, CA, USA: International Journal of Information Management, vol. 68, Mar. 2022, Morgan Kaufmann Publishers Inc., 1998, pp. 580–592. Art. no. 102496. https://doi.org/10.1016/j.ijinfomgt.2022.102496 [54] M. A. Mahdi, K. M. Hosny, and I. Elhenawy, “FR-Tree: A novel rare [67] D. Di Fatta, D. Patton, and G. Viglia, “The determinants of conversion association rule for big data problem,” Expert Systems with Applications, rates in SME e-commerce websites,” Journal of Retailing and Consumer vol. 187, Jan. 2022, Art. no. 115898. Services, vol. 41, pp. 161–168, Mar. 2018. https://doi.org/10.1016/j.eswa.2021.115898 https://doi.org/10.1016/j.jretconser.2017.12.008 [55] T. Kim, D. Kim, and Y. Ahn, “Instant customer base analysis in the financial services sector,” Expert Systems with Applications, vol. 202, Doae Mensouri received the Diploma degree in Computer System and Apr. 2022, Art. no. 117326. https://doi.org/10.1016/j.eswa.2022.117326 Decision Support from the University of Science and Technology UAE [56] K. Jerath, P. S. Fader, and B. G. S. Hardie, “New perspectives on customer Morocco, in 2018. She is in her final year as a PhD student at the UAE ‘death’ using a generalization of the Pareto/NBD model,” 2011. [Online]. University, Tangier, Morocco. She has a publication in the International Journal Available: of Advanced Computer Science and Applications 2022, https://papers.ssrn.com/sol3/papers.cfm?abstract_id=995558 HTTPS://DOI.ORG/10.14569/ijacsa.2022.0130658. [57] X. Chen, R. Geng, and S. Cai, “Predicting microblog users’ lifetime Her research interests include IT, marketing, business intelligence, artificial activities – A user-based analysis,” Electronic Commerce Research and intelligence, decision making, and customer relations. Applications, vol. 14, no. 3, pp. 150–168, May 2015. E-mail: doae.mensouri@etu.uae.ac.ma https://doi.org/10.1016/j.elerap.2014.06.001 ORCID iD: https://orcid.org/0000-0003-2082-2549 [58] J. Dipak and S. S. Siddhartha, “Customer lifetime value research in marketing: A review and future directions,” 2002. [Online]. Available: Abdellah Azmani received his PhD degree in Industrial Computing in https://www.proquest.com/openview/1f7fc325cb76ec2ede9989ff9defb1 Dynamic System Modelling and Artificial Intelligence at the University of 77/1?pq-origsite=gscholar&cbl=32002. Accessed on: Dec. 06, 2021. Science and Technology of Lille in 1991. He worked as a Professor at the Ecole [59] T. Reutterer, M. Platzer, and N. Schröder, “Leveraging purchase Centrale de Lille, France and at the Institute of Computer and Industrial regularity for predicting customer behavior the easy way,” International Engineering from Lens, France. He is a Professor at the Faculty of Science and Journal of Research in Marketing, vol. 38, no. 1, pp. 194–215, Mar. 2021. Technology of Tangier, Morocco. He is a member of the Laboratory of https://doi.org/10.1016/j.ijresmar.2020.09.002 Informatics, System and Telecommunication (LIST) and he has established the [60] R. Colombo and W. Jiang, “A stochastic RFM model,” Journal of Intelligent Automation team, which he coordinates. He has contributed to many Interactive Marketing, vol. 13, no. 3, pp. 2–12, Aug. 1999. theses, scientific research projects, as well as he elaborates and produces many https://doi.org/10.1002/(SICI)1520-6653(199922)13:3<2::AID- IT and decision support solutions for public administration, business DIR1>3.0.CO;2-H management, marketing and logistics. [61] P. S. Fader and B. G. S. Hardie, “The Gamma-Gamma model of monetary E-mail: a.azmani@uae.ac.m value,” 2013. [Online]. Available: http://brucehardie.com/notes/025/

Journal

Applied Computer Systemsde Gruyter

Published: Dec 1, 2022

Keywords: Apriori algorithm; BG/NBD Gamma-Gamma CLV & RFMTS models; marketing recommendation system; smart offer generation

There are no references for this article.