Contextual-Boosted Deep Neural Collaborative Filtering Model for Interpretable Recommendation

Abstract Collaborative filtering (CF) is one of the most successful recommendation techniques due to its simplicity and attractive accuracy. However, existing CF methods fail to interpret the reasons why they recommend a new item. In this paper, we propose a Contextual-boosted Deep Neural Collaborative filtering (CDNC) model for item recommendation, which simultaneously exploits both item introductions (textual features) and user ratings (collaborative features) to alleviate the cold-start problem and provide interpretable item recommendation. Specifically, we propose an interactive attention mechanism to learn the user representation, which makes use of the mutual information from both the user ratings and item introductions to supervise the representation learning of each other. With the learned attention weights, we can obtain the importance of each historical item among the historical list. Meanwhile, the attention model can assign different weights to the words in item introductions according to their importance. Therefore, CDNC can provide interpretations for the recommendations by assigning different attention weights to the historically interacted items and the words in the item introductions. On the other hand, we also learn the distributed representations of new-coming items with deep neural networks (i.e., LSTM), considering both rating and item introduction information. Finally, the user representation and the representations of new-coming item are concatenated to perform recommendation score prediction. Extensive experiments on four public benchmarks demonstrate the effectiveness of CDNC. In addition, CDNC has the advantage of interpreting the recommendations and providing user profiles for down-stream applications.