Title:Convolutional Quantum-Like Language Model with Mutual-Attention for Product Rating Prediction

Abstract:Recommender systems are designed to help mitigate information overload users experience during online shopping. Recent work explores neural language models to learn user and item representations from user reviews and combines such representations with rating information. Most existing convolutional-based neural models take pooling immediately after convolution and loses the interaction information between the latent dimension of convolutional feature vectors along the way. Moreover, these models usually take all feature vectors at higher levels as equal and do not take into consideration that some features are more relevant to this specific user-item context. To bridge these gaps, this paper proposes a convolutional quantum-like language model with mutual-attention for rating prediction (ConQAR). By introducing a quantum-like density matrix layer, interactions between latent dimensions of convolutional feature vectors are well captured. With the attention weights learned from the mutual-attention layer, final representations of a user and an item absorb information from both itself and its counterparts for making rating prediction. Experiments on two large datasets show that our model outperforms multiple state-of-the-art CNN-based models. We also perform an ablation test to analyze the independent effects of the two components of our model. Moreover, we conduct a case study and present visualizations of the quantum probabilistic distributions in one user and one item review document to show that the learned distributions capture meaningful information about this user and item, and can be potentially used as textual profiling of the user and item.