Shaping the causes of product returns: topic modeling on online customer reviews

Abstract Product return is a common phenomenon in the online retailing industry and entails several inconveniences for both the seller, who incurs in high costs for restocking the returned goods, and the customer, who has to deal with product re-shipping. In this paper, we outline a data-driven approach, based on Natural Language Processing, in which a broad corpus of customer reviews of an online retailer is exploited with the aim of shaping the main causes of product returns. In particular, a variety of topic modeling techniques represented both by classic methods, given by LDA and variants, and more recent algorithms, i.e., BERTopic, were applied to identify the main return reasons across multiple product categories, and their outcomes were compared to select the best approach. The category-dependent sets of return causes inferred through topic modeling largely enrich the product-agnostic list of return reasons currently used on the e-commerce platform, and provide valuable information to the retailer who can devise ad-hoc strategies to mitigate the returns and, hence, the costs of the related logistic network.