Research On Collaborative Filtering With GNN Representation Learning

As an efficient information filtering tool,recommender system can effectively alleviate the problem of "information overload".Collaborative filtering is the most widely used algorithm in recommender systems.It analyzes user preferences through users’ historical behaviors,models user characteristics,and recommends items of interest to users.For collaborative filtering needs to use the user’s historical behavior data,and such data is often small relative to the number of items,collaborative filtering content recommendation faces serious data sparsity problems.Recently,recommendation based on graph neural network has received more and more attention because the association between users and items can form a natural bipartite graph.The user-item bipartite graph and graph convolution methods can effectively enrich the information of target nodes and alleviate data sparsity.However,there are still problems in the recommendation based on graph network.First of all,the graph convolution model is too complex,and it cannot fully play the role of its feature transformation and nonlinear activation in the recommendation task to tap potential correlations.Secondly,when aggregating node information,equal looking at various relationships,merging them directly with the target node information without processing results in unreasonable expression of the final target features.Thirdly,the current message aggregation method of graph convolution is explicit addition or splicing,which leads to the model training efficiency suffers,and training becomes very slow when the dataset is large.Another way to alleviate data sparseness is to introduce auxiliary information,such as user behavior information,timestamps or user reviews,item introductions,etc.However,the current recommendation model using behavior information only mines the relationship between different behaviors and ignores the relationship between users.The connection of behaviors results in that the behavioral information cannot fully describe the user’s interest.The model based on the graph network is also subject to the explicit message passing algorithm,and the model training efficiency is not high.Aiming at the problems existing in the current collaborative filtering recommendation algorithm based on neural network,the research contents of this paper are as follows:(1)In traditional collaborative filtering models,the feature vector generated by one-hot encoding is sparsely informative.Heterogeneous behavior data is only employed to describe the relationship between different behaviors and the relationship between behaviors of different users is ignored.Aiming at these problems,an algorithm of collaborative filtering with heterogeneous neighborhood aggregation is proposed.Firstly,the heterogeneous interaction between users and items is modeled by the graph,and neighborhoods are built through the connectivity of graph.Then,the neighborhood information integrated by the lightweight graph convolution method is merged into the feature vectors of the target users and items.Finally,the feature vectors of users and items integrating with neighborhood information are input into a multi-task heterogeneous network for training.The problem of data sparseness is alleviated by enriching the hidden information of feature vectors.Experiments on the datasets prove that the performance of the proposed model is better.(2)The message passing mechanism adopted for the recommendation algorithm based on graph convolution is inefficient,which leads to the model spending a lot of time training and traditional recommendation models usually only use one type of user behavior data to train the model,and only one type of user behavior cannot fully express user interests.This paper proposes an algorithm named collaborative filtering fusing implicit neighborhood associations and heterogeneous behavioral characteristics.First of all,the heterogeneous interactions between users and items are modeled through graphs,and neighborhoods are constructed through the connected properties of the graphs;Secondly,through the implicit neighborhood aggregation algorithm,the effective information of higher-order neighborhoods can be fused into the target feature vector,and the cost of training can be reduced at the same time;Then,weights are assigned to different interactions through implicit item interactions,which leads to the more accurately description to the importance of different associations to user or item;Finally,the heterogeneous features are input into the multi-task heterogeneous network,and the whole model is optimized by fusing the multi-task loss based on non-sampling and the neighborhood loss based on implicit graph information messaging.Experiments are carried out on Movielen-1m,Beibei,Taobao three data sets,HR and NDCG indicators are used to evaluate the model,the results obtained by this model are better than other recommendation models of the same type,and the model training time is shortened and the efficiency is improved.