Improving an e-commerce’s recommendations accuracy by exploiting impression data.

Abstract

Recommender systems serve the purpose of recommending items to users in online environments such as a streaming service, an E-commerce, a news aggregator. As both computational and networking capabilities significantly improved over the years, the volumes of user activity data that are collected exponentially grew as well, providing intelligent systems like modern recommender systems with additional, precious sources of information. An example of voluminous and potentially valuable kind of data is given by impressions, which are tracked records of users simply being displayed items on online pages.

In this work we will perform an analysis of impressions that occurs repeatedly between the same user-item couples, on an impression dataset from a luxury fashion e-commerce. We will observe interesting behavioral patterns, which suggest that users that views items multiple times interacts more frequently with them, and we will try to model such patterns. Finally, we will demonstrate the potential value of this source of information by incorporating it offline in the existing recommendation pipeline, implementing a plug-in weighting model that, based on previous repeated impressions, boosts or penalizes the recommendation score given to items by the the recommendation model in place. We will see that it was possible to increase the recommendation accuracy without giving up diversity, proving that in this context impressions can represent a valid and informative input to the recommender system.