Identification and Utilization of Contextual Features Using Post-Filtering in an E-Commerce Context-Aware Recommender System

Abstract

Recommender systems are systems that provide recommendations to a user based on information gathered about that user or by finding other similar users. Only a few years ago, these recommender systems were solely dependent on either explicit information given by the user or implicitly gathered information such as user patterns- and actions. During these last few years we have seen a slightly increased interest incorporating context within recommender systems in order to provide better recommendations. There have been however, little or no contribution to find those contextual features that are useful and relevant to recommender systems that operates in an e-commerce setting. By exploiting more of the information that lies in the users’ context, it’s possible to generate more personalized recommendations related to the context. For instance, if a person at work visits an online store, the person might prefer getting work-related recommendations compared to personal-related. Applying contextual features into recommender systems could convert window-shoppers into buyers, increase cross-sell and customer satisfaction and ultimately generate more revenues to the company. In addition, the consequences of using the proposed contextual features in a fine-grained collection compared to a coarse-grained one will be studied and analyzed. To start on the task of finding relevant contextual features and how they can affect the final recommendation outcome, a survey was conducted in companionship with one of Norway’s biggest auction-based online stores, where over 35000 customers participated.