Both reviews and user-item interactions (i.e., rating scores) have beenwidely adopted for user rating prediction. However, these existing techniquesmainly extract the latent representations for users and items in an independentand static manner. That is, a single static feature vector is derived to encodeher preference without considering the particular characteristics of eachcandidate item. We argue that this static encoding scheme is difficult to fullycapture the users' preference. In this paper, we propose a novel context-awareuser-item representation learning model for rating prediction, named CARL.Namely, CARL derives a joint representation for a given user-item pair based ontheir individual latent features and latent feature interactions. Then, CARLadopts Factorization Machines to further model higher-order featureinteractions on the basis of the user-item pair for rating prediction.Specifically, two separate learning components are devised in CARL to exploitreview data and interaction data respectively: review-based feature learningand interaction-based feature learning. In review-based learning component,with convolution operations and attention mechanism, the relevant features fora user-item pair are extracted by jointly considering their correspondingreviews. However, these features are only review-driven and may not becomprehensive. Hence, interaction-based learning component further extractscomplementary features from interaction data alone, also on the basis ofuser-item pairs. The final rating score is then derived with a dynamic linearfusion mechanism. Experiments on five real-world datasets show that CARLachieves significantly better rating prediction accuracy than existingstate-of-the-art alternatives. Also, with attention mechanism, we show that therelevant information in reviews can be highlighted to interpret the ratingprediction.
展开▼