Post-Click Behaviors Enhanced Recommendation System

Abstract

To predict users’ interests, the traditional recommendation system (RS) relies on exploring the explicit user-item ratings and macro implicit feedbacks (e.g., whether or not a user clicks the item). In this work, fine-grained post-click behaviors (e.g., mouse behaviors, keyboard events, and page scrolling events) are integrated to alleviate the data sparsity problem of explicit feedback and the data accuracy problem of macro implicit feedback. In the deployed article recommendation pipeline, a variety of post-click behaviors are combined to create a reading pattern model. The reading patterns are leveraged by the recommendation system to estimate users’ preference levels. As compared with existing click-based (macro implicit feedback) and dwell time-based (single micro implicit feedback) recommendation systems, the test performance of our designed reading pattern-based RS has been significantly improved in terms of rating prediction and ranking.