TCR: Temporal-CNN for Reviews Based Recommendation System

In recent year, it has become a popular trend that online stores encouraged their users to write review texts for shopping items. Obviously, these collected text-reviews are helpful for understanding item properties and user preferences, as well as improving the quality of recommendation. However, existing works put considerable attentions on the performance of recommendation without using the temporal information, while the customer inclinations are evolving. In this paper, we propose TCR to model user preferences and item properties by using the convolutional neural network (CNN) combined with temporal information. In details, since the item popularity and user preferences are constantly evolving, we then build a time model that to capture the influence of time evolving on the performance of recommendation and integrate the proposed time model to the original CNN recommender. Furthermore, aiming at building an effective model, we carry out the experimental analysis on the influence of four factors (i.e., word vector embedding dimension, word frequency of comment text, the depth and width of CNN model) on the performance of recommender system. Based on the theoretical analysis, we identify the key factors, and use these factors to optimize our TCR model. Finally, we conduct the experiments on the industrial dataset, i.e., Amazon. It demonstrates that our proposed model has achieved better results than the existing models in terms of prediction accuracy.