Deep Recommendation Model Based on Local Attention and GRU

The recommendation systems are dedicated to providing people with valuable information by information filtering. Although the recommendation models based on matrix decomposition are widely used, they perform poorly in the case of sparse data. The review-based recommendation models use reviews to extract user preferences and item features which can alleviate the data sparsity problem, but the improvement of recommendation quality is limited because these models didn't consider the weights of words or phrases in the reviews. This paper proposes a Deep recommendation model based on Local Attention and GRU (DLAG) for reviews level explanation. In DLAG, we first use the local attention to assign the weight to a word or phrase in the reviews. Then, the convolutional neural networks (CNNs) extract complex features after local attention. Finally, the bidirectional gated recurrent unit (Bi-GRU) can extract deep levels non-linear features from user's reviews and item's reviews, and generate latent hidden vectors. Multi-layer perceptron training is used to predict user's rating for an item. This model alleviates the problem of data sparseness by weighting words in reviews and extracting serializable features. Through training on three different datasets, we found that DLAG achieves better performances than others models.