Trust-Aware Collaborative Filtering with a Denoising Autoencoder

摘要

Collaborative filtering is one of the most successful and extensive methods used by recommender systems for predicting the preferences of users. However, traditional collaborative filtering only uses rating information to model the user, the data sparsity problem and the cold start problem will severely reduce the recommendation performance. To overcome these problems, we propose two neural network models to improve recommendations. The first one called TDAE uses a denoising autoencoder to integrate the ratings and the explicit trust relationships between users in the social networks in order to model the preferences of users more accurately. However, the explicit trust information is very sparse, which limits the performance of this model. Therefore, we propose a second method called TDAE++ for extracting the implicit trust relationships between users with similarity measures, where we employ both the explicit and implicit trust information together to improve the quality of recommendations. Finally, we inject the trust information into both the input and the hidden layer in order to fuse these two types of different information to learn more reliable semantic representations of users. Comprehensive experiments based on three popular data sets verify that our proposed models perform better than other state-of-the-art approaches in common recommendation tasks.