Explainability in recommender systems: a comparative analysis of two transformer-based models

In e-commerce, recommender systems have long been important for improving user experience and boosting sales conversion rates. However, as these systems become increasingly complex, it becomes harder to understand the reasoning behind the recommendations. This project will carry out an explainability analysis of two models of recommender systems: a transformer-based model used by the company Decathlon for website personalized recommendations and a two-tower-based one (referred to as the challenger model) developed from scratch for the same purpose. The explainability analysis will focus on specific parts of the models, namely the embeddings learned at different stages, the attention layers present in the transformer modules of both the production and the challenger model, and the final embedding space. This analysis will try to shed light on the influence that different external factors such as price, sales, or seasonality have on the latent representation learned by the model. Even though the challenger model does not surpass the production model in accuracy metrics, it shows great strength in its exploration rate and its coverage of the items catalog, likely due to how it addresses the cold-start problem. Moreover, the explainability analysis evidences how the models excel at learning information about the products and users that are not explicitly fed to them as inputs.