Attentive convolutional neural network with the representation of document and sentence for rating prediction

Guisheng Yin, Fukun Chen, Yuxin Dong, Gesu Li

Abstract: With the development of the internet and the exploding data, the recommendation algorithm is an effective way to solve information overload has been widely concerned. Recently, some researchers have introduced review information to improve recommendation performance. Previous methods typically combine the reviews for a given user or item into a single long document and then extract features from the document level. However, such methods train the model on aggregated historical reviews over time. There may be many conflicting semantics or vocabularies, so scalability and accuracy may be affected. This paper proposes an attentive convolutional neural network representing document and sentence(ACNNDS) for rating prediction. The parallel network includes the sentence-level and document-level reviews, which extract user and item latent expressions from aggregated documents. The importance of sentences and words in reviews is different for different users. We established sentence-level and document-level attention mechanisms to capture the critical sentences and words from reviews to solve this problem. The experimental results show that the ACNNDS boosts recommendation performance, which is significantly better than comparing methods.