Customer Behavior Analysis for Marketing Applications

Abstract

In recent years, we have seen enormous growth in retail activity [45], and consequently, the companies register a large amount of transactional data about products bought by their customers in different visits (orders, transactions). Implicitly, transactional data provides meaningful information about Customer Purchasing Behaviors (CPBs) in a company. CPB refers to the actions taken by customers, such as customer purchase decisions, preferences and interests, repetitions or order frequency, compositions of promotional items inside each order, products that are more likely to be bought together or buy frequently, etc. Data Mining (DM) techniques offer huge opportunities for extracting insights about hidden CPB patterns from transactional data. To cope with these issues, there are two main data mining approaches. Clustering attempts to discover customer segments based on similar buying patterns helping businesses better understand customers’ preferences, and facilitating the development of customized new products and product offerings. Meanwhile, recommendation aims to predict customer preferences and recommend items that may be most suitable for them based on their previous behaviors. However, discovering, extracting, and predicting knowledge and behavioral patterns are not easy tasks. Because different dimensions come to play (e.g., customers, products, orders, and categories) the number of elements related to dimensions are often extensive; this high dimensionality and the sparsity of data could be a source of poor clustering and recommendation algorithm results. Consequently, the crucial questions concerning the quality of these models are: How to transform the transactional data and quantify the interactions between customers comprising all involved dimensions? How to accomplish the problem of behavior prediction through a more expanded set of categories? How to represent, in a more interpretable way, the interactions between customer buying preferences? The customer complex network definition through objects of the Multiple Instance (MI) space is proposed in this thesis. Mapping the transactional data into the correct data transformation and representation model is essential to improve the algorithms that discover and predict buying preferences accomplishing the application domain restrictions in the CPB analysis. The developed segmentation methodology and recommendation for CPB analysis tackle the problem from different directions. First, different customer network construction models are proposed, analyzed, and compared through empirical experiments to discover customer communities based on the application domain restrictions — compactness and granularity level. Second, a sequence prediction algorithm is proposed, which is analyzed and compared with other classical approaches, mainly focuses on estimating the next buy preference capabilities through a more expanded set of previous preferences- rather limited in existing approaches. Third, the proposed Rough Net definition contributes to the customer segmentation analysis from three perspectives: the validation, the evolutionary estimation in dynamic customer networks, and the visualization in the interpretable way of customer community (segment) interactions. The experiments in this thesis show that different customer segmentation can be obtained mainly depending on business perspectives and domain restrictions. The proposed Customer as a Bag of Orders (CBO) transformation and MInteraction similarity allow obtaining more compactness customer segments (communities). The developed two-stage multi-category segmentation approach allows deriving a more stable granularity level. It is useful for real domains with high dimensional features where it is desired to expand the customer preferences through cross-selling and upselling campaigns. The proposed deep learning sequence prediction approach, named DeepCBPP, allows to successfully learn on transactional data the temporal dependencies of preferences in scenarios where typically the orders are small, and the purchase frequency and the number of orders are low. In addition, to be focused on the enterprise perspectives, both CPB approaches are also applied to two study cases — GameMania company and Spanish supermarket. Both study cases show that the obtained results have added value, which can help to improve the marketing applications under consideration. Our proposal allows analyzing the interaction and evolution of — multiplex or monoplex — customer complex networks for different marketing applications. From an operational perspective, the derived and integrated results would help retailers better understanding the behavioral patterns of customers, establishing effective and customized actions to retain more customers, maximizing boost product revenue, improving customer experience, and optimizing the store design.