Abstract: To fully combine the features of review text in recommendation systems, a recommendation model fully combining review features based on dual attention (FCRF) is proposed. First, the model encodes the review text by utilizing the encoder model, which is fine-tuned by the sentiment classification task to obtain the corresponding feature embedding. Secondly, to realize the cross combination between user features and item features, the bilinear inner product is adopted to calculate the cross attention of review feature embeddings between the user and the item. Thirdly, to obtain the final feature representation of the user and the item, the multi-head self-attention is adopted to realize the self-combination of the user's review features and the item's review features. Experimental results on four real-world datasets show that the mean square error of FCRF is 1.43% lower than that of the state-of-the-art model, which verifies the effectiveness of FCRF.

Key words: feature combination, encoder model, bilinear inner product attention, multi-head self-attention