Context Information in a Collaborative Recommender System Deployed in Real Environment

Abstract

Modern e-commerce solutions (WWW services, e-stores, news portals) develop continuously, gathering and offering more and more new, interesting and various items. Unfortunately, common users are not able to deal with this information overload and reach most of them. They limit to the most popular, however often not the most interesting to them resources. A solution for this problem are personalized recommender systems. There are some popular and effective methods to build a good recommendation system: collaborative filtering, content-based, knowledge-based and hybrid. Another approach, which made a significant progress over the last several years, are context-aware recommenders. There are many additional information related to the context or application area of recommender systems, which can be useful to generate accurate propositions, e.g. user localisation, items categories or attributes, a day of a week or time of a day, weather. Important issue is evaluation of recommender systems effectiveness. Usually, they are only assessed with respect to their prediction accuracy (RMSE, MAE). However, in real environment recommendation lists are finally evaluated by users who take into consideration many various factors, like novelty or diversity of items. In this article a multi-module collaborative filtering recommender system with consideration of context information is presented. The context is included both in post-filtering module as well as in similarity measures simply extended with category relationship. Evaluation was made off-line with respect to prediction accuracy and on-line, on real shopping platform.