PerceptRank: A Real-Time Learning to Rank Recommender System for Online Interactive Platforms

In highly interactive platforms with continuous and frequent content creation and obsolescence, other factors besides relevance may alter users’ perceptions and choices. Besides, making personalized recommendations in these application domains imposes new challenges when compared to classic recommendation use cases. In fact, the required recommendation approaches should be able to ingest and process continuous streams of data online, at scale and with low latency while making context dependent dynamic suggestions. In this work, we propose a generic approach to deal jointly with scalability, real-time and cold start problems in highly interactive online platforms. The approach is based on several consumer decision-making theories to infer users’ preferences. In addition, it tackles the recommendation problem as a learning-to-rank problem that exploits a heterogeneous information graph to estimate users’ perceived value towards items. Although the approach is addressed to streaming environments, it has been validated in both offline batch and online streaming scenarios. The first evaluation has been carried out using the MovieLens dataset and the latter targeted the news recommendation domain using a high-velocity stream of usage data collected by a marketing company from several large scale online news portals. Experiments show that our proposition meets real world production environments constraints while delivering accurate suggestions and outperforming several state-of-the-art approaches.