A Deep Neural Network Model for Rating Prediction Based on Multi-layer Prediction and Multi-granularity Latent Feature Vectors

Recommender systems have attracted abundant research in the past decades. Side information is generally used besides the rating matrix to alleviate the data sparsity problem for recommendation models. To achieve better performance, in recent years deep learning (DL) technique has been introduced to recommendation models. It can be noted that most existing recommendation models incorporating DL technique only use one layer as the learned features; and the learned features for all users/items have the same dimension despite the fact that different users/items have different numbers of ratings. The aforementioned issues have negative impact on the performance of these recommendation models. To address the issues, in this paper we propose a Deep neural network model based on Multi-layer prediction and Multi-granularity latent feature vectors (DMM model). The DMM model has two features: (1) A user or an item is represented by multiple latent vectors with different granularity, which can better describe the relationships between users and items. (2) Each layer in the DMM model produces a predicted rating for given user and item, then the overall rating is calculated by combining all these predicted values, which ensures fully use of the information in rating matrix and side information and thus may result in better performance. Experimental results on three widely used datasets demonstrate that the proposed DMM model outperforms the compared models.