Research On Recommendation Algorithms Based On Graph Convolutional Network And Neural Collaborative Filtering

Recommender Systems(RS)are one of personalized information filtering tools that can effectively alleviate the information overload problem.Top-N recommendation is one of the main problems in recommender systems.And its goal is to sort the candidate items based on the target user's historical behaviors(such as a view,click,collect and purchase,etc.)and other related datas and then generate a ranked list of size-N items for the target user.To address the top-N recommendation problem,the existing collaborative filtering algorithms based on graph convolutional networks focus on representation learning and often ignore the learning of matching function between users and items.They usually apply simple and fixed dot product to model complex interaction between users and items.Meanwhile,the existing ones based on neural collaborative filtering focus on matching function learning and often neglect the learning of the feature vectors for users and items.They usually use a linear embedding layer to learn the user and item representations.In order to overcome the shortcomings of the above two types of methods,this paper makes the following contributions:First,this paper proposes an improved collaborative filtering algorithm named Graph Convolutional Matrix Factorization(GCMF)and its variant named GCMF-n based on graph convolutional networks.These algorithms consider that different graph convolutional layers contain different semantic informations,which are of different importance to different users and items.Therefore,GCMF and GCMF-n propose to learn the aggregation weights of different layers for each user and item.Second,this paper further devises a graph convolutional collaborative filtering algorithm named GCMF-a based on GCMF-n and attention mechanism.The algorithm considers that the feature vector of each graph convolutional layer is a feature of users or items.When modeling the interaction between users and items,different user-item feature pairs have different importance.Therefore,it proposes to utilize the attention mechanism to adaptively assign weights to different user-item feature pairs.And the experimental results demostrate that GCMF-a provides better recommendation performance compared with the state-of-the-art methods.Third,this paper presents a general recommendation framework named GCN-NCF,short for Graph Convolutional Network based Neural Collaborative Filtering.The proposed framework combines the strengths of representation learning-based and matching function learning-based methods,which first adopts graph convolutional networks to learn the latent vectors for users or items and then employs neural collaborative filtering framework to learn the matching function between users and items.The instantiations of GCN-NCF achieve superior performance over other state-of-the-art algorithms.