MARF: User-Item Mutual Aware Representation with Feedback

As deep learning (DL) technologies have developed rapidly, many new techniques have become available for recommender systems. Yet, there is very little research addressing how users' feedback for particular items (such as ratings) can affect recommendations. This feedback can assist in building more fine-grained user profiles, as not all raw clicks will truly reflect a user's preference. The challenge of encoding such records, which are typically prohibitively long, also prevents research from considering using the whole click history to learn representations. To address these challenges, we propose MARF, a novel model for click prediction. Specifically, we construct fine-grained user representations (by considering both the multiple items browsed, and user's feedback on them) and item representations (by considering browsing histories from multiple users, and their feedback). Moreover, the flexible up-down strategy is designed to avoid loading incomplete or overloaded historical information by selecting representative users/items based on their feedback records. A comprehensive evaluation on three large scale real-world benchmark datasets, showing that MARF significantly outperforms a variety of state-of-the-art solutions. Furthermore, MARF model is evaluated through an ablation study that validates the contribution of each component. As a final demonstration, we show how MARF can be used for cross-domain recommendation.