Research On Collaborative Filtering With Convolutional Neural Networks

In the era of information explosion,recommender system has been widely used for various online services to alleviate information overload.As a main technique for recommender system,collaborative filtering(CF)infers user preference not only from the history behavior data of the user but also from the behavior data of others.In recent years,neural network has achieved excellent performance in many research fields such as computer vison,natural language processing,so it gains popularity among recommender systems.Neural Collaborative Filtering algorithm concatenates user-item embedding and utilizes multiple-layer perceptron to learn user-item interaction relationships achieving good performance on implicit feedback recommendation task.Yet later work proves that such embedding concatenation method cannot express the complicated structure relations over user-item embedding.In order to learn more accurate dimension correlation of user-item embedding,we propose to use stacked embedding structure to combine user-item embeddings and employ convolutional neural network to learn multiple aspects of dimension correlations for user-item embeddings.On the one hand,stacked embedding makes user-item embedding dimension present local relations so that we can build better local dimension correlation comparing with embedding concatenation.On the other hand,stacked embedding structure can easily be extended with transformed user-item embedding or latent factors.Convolutional neural network is so good at capturing local features that we can employ it to effectively mining local relationships of embedding dimension.We adopt public accessible datasets of Amazon to conduct experiments for the proposed convolutional neural collaborative filtering with stacked embedding method and compare it with the prevailing neural network based collaborative filtering methods on Top-K recommendation task.The experiments show that the proposed convolutional neural collaborative filtering method achieves significant improvement over the state-of-the-art neural network based collaborative filtering methods.