PER: A Probabilistic Attentional Model for Personalized Text Recommendations

In many recommendation domains, items to be recommended are associated with text. We observe that for an item, customers are usually attracted by parts of its associated text rather than the whole one. For example, a researcher may decide to read a paper if some of its words or sentences are matched with his or her own interests. However, previous methods fail to attentively focus on different parts of text according to users' personal interests. In this paper, we first introduce a novel Personalized Attentional Network (PAN) to capture parts of text matched with a user's personal interests. The network is able to adapt to a user's personal interests and capture relevant parts of text for the user. Then, we propose a probabilistic attentional model for PErsonalized text Recommendation (PER). PER further integrates PAN into a probabilistic framework, which leads to a better generalization. In the experiments, we validate the effectiveness of the proposed model (PER) and show that on average, PER improves the strongest baseline by 18.2% and 14.2% in terms of Recall and Mean Average Precision (MAP), respectively.