A hybrid recommendation method that combines forgotten items and non-content attributes

Abstract

Recommender Systems (RSs) play important role in many Web applications nowadays, helping users to find their favorite items amid a huge number of options. Among numerous open challenges inherent to RSs, this dissertation addressed the challenge of enhancing the discovery of potentially relevant items for each user. In this sense, we exploited two algorithmic limitations unaddressed in the literature. First, RSs fail to bring back items consumed long ago that are potentially relevant for users nowadays. Second, RSs fail to capture the whole extent on which implicit signals of preferences observed on past consumption relate to preferences observed on current consumption. We addressed the first limitation by reviewing the users long-term history and identifying a subset of consumed items forgotten but still re-consumable (i.e., forgotten re-consumable items). We mitigated the second limitation by explicitly modeling the subset of attributes derived from metadata or consumption data (i.e., non-content attributes). Finally, we proposed ForNonContent, a hybrid method that addresses both limitations simultaneously. Besides validating these algorithmic limitations, offline analysis on four real datasets demonstrated that recommending forgotten re-consumable items may bring diversified and novel recommendations. Also, we found that non-content attributes may enhance ecommendations of six major RSs. Furthermore, we identified a complementary nature of the enhancements associated to each limitation. Finally, a user study with MovieLens users demonstrated that they valued more the recommendations issued by ForNonContent. In summary, this work pointed out a new and promising direction to enhance the user experience with RSs.