Predicting l-CrossSold products using connected components: A clustering-based recommendation system

Abstract

• We provide a novel data representation for the point-of-sale transactional datasets. • We have developed a novel similarity measure that captures the product’s context. • A novel algorithm to predict the cross-sold items. • Our algorithm obtains the top N recommended items for an active user efficiently. • Our algorithm retrieves the most frequent items with a significant performan. Recommender systems (RSs) are an integral part of the online retail industry. They create tremendous potentials for cross-selling, including increasing sales revenue, improving consumer fulfillment, increasing customer lifetime value, controlling product consumption to optimize resources, and dynamically adapting to consumer behaviors. Online retailers have relied on the position and presence of RSs to assist users in better decision-making and thereby increase their revenues. However, with dynamic changes in consumer behavior patterns, the quality and accuracy of recommendations have become a significant challenge for e-commerce retailers. Moreover, with the rapid evolution of online data, most RSs suffer from data sparsity and scalability problems. This paper introduces a new model of applying clustering analysis concepts to RSs along with advancements in graph theory to react effectively to user changes and business challenges in the the online retail industry. A clustering-based RS using the notion of the cross-sold score—namely, “ l - CrossSold ”—is proposed and tested with the aim of enhancing the quality of recommendations, handling data sparsity, improving consumer profiling and addressing scalability in the current recommendation methods in order to generate a more practical set of personalized recommendations. The proposed algorithm shows a significant improvement in both the prediction accuracy and the speedup as compared to the state-of-art collaborative filtering RSs, clustering-based collaborative filtering RSs, and rule-based RSs.