Learning to recommend similar alternative products in e-Commerce catalogs

Abstract

In this work, we describe a novel method we designed, implemented and tested to finding products that are similar alternatives to a given product in the catalog of an e-commerce site. By similar alternatives, we mean products that, although are not identical to a product of interest, have features that make them suitable alternatives for customers that look for it. Our motivation is to enable the recommendation of alternativeproductsbasedsolelyontheproduct’sfeatures,withoutrelyingonhistorical purchase data. By doing so, we address the so-called cold start problem, which is often found in product recommendation approaches, and that may lead to profit loss in ecommerce sites. Our method, we call GPClerk, uses Genetic Programming (GP) to learn functions for comparing two products and telling whether two products are similar alternatives or not. These functions are termed here as product comparison functions. To make our method feasible in typical e-commerce settings, we also propose an unsupervised strategy to generate training examples to be used in the learning process. Results of experiments we carried out and report here indicate that our method is capable of generating suitable product comparison functions and that our strategy for automatically generating training data is effective for this task.