Pruning Graph-based Recommendation Models Based On Lottery Ticket Hypothesis

The recommendation system is an information filtering tool which aims to recommend items to the appropriate user.It can effectively solve the problem of information overload in the era of big data and build a win-win bridge for information producers and consumers.In the field of recommendation system,graph-based model has been widely used with its powerful representation capability.The graph recommendation model generally constructs user-item interaction records into a bipartite graph,and then mines the interaction information of user-item nodes through graph convolution operation to finally provide accurate personalized recommendation service.Thanks to the rapid development of deep learning technology in recent years,the recommendation accuracy of the graph recommendation models has been continuously improved.However,this paper finds that the existing graph recommendation models have shortcomings in terms of efficiency and effectiveness:1)The number of user-item representation matrix parameters is very large,and the dimension of representation vector is the same to users and items,which makes it difficult to fully represent the heterogeneity among different users and items.2)The scale of the user-item interaction graph is often very large,making graph convolution operations consume a large amount of computational resources and are susceptible to the influence of noisy edges in the interaction graph.To solve these problems,we study the lottery hypothesis theory for graph recommendation models,and lighten the representation matrix and interaction graph of graph recommendation models,respectively.In this paper,the two major shortcomings of graph recommendation models are discussed as follows:1.We propose a pruning method for the representation matrix of graph recommendation model,in which the parameters of the representation matrix are pruned in an iterative way to break the limitation of consistent representation dimensionality while reducing the parameter scale.This method not only finds sparse representation matrixes(the winning lottery tickets)comparable to the representation capability of the original dense representation matrix,but also demonstrates that the winning lottery tickets are widely existing in the representation matrix of the graph recommendation model.In addition,compared to the original representation matrix,the winning lottery tickets have significant advantages in time,space and recommendation accuracy.On the three datasets of Yelp2018,TikTok and Kwai,the winning lottery tickets only use 29%～48%,7%～10%,3%～17%of the parameters,respectively,which can achieve the test performance similar to the dense representation matrix.2.We propose a pruning method for the interaction graph structure of graph recommendation model to mitigate the effect of noisy edges while reducing the scale of the graph structure by iteratively pruning the interaction graph.This method is an improvement of the method in 1.By a random rewinding mechanism,the low fault tolerance problem of the interaction graph pruning is overcome,and the stability and effectiveness of the algorithm is improved.Experimental results show that this method can stably find sparse sub-interaction graphs(the winning lottery tickets)with close performance to the original interaction graph.In addition,compared with the original interaction graph,the winning tickets can greatly reduce the computational complexity of the model.On the three datasets of Yelp2018,TikTok and Kwai,the winning lottery tickets the computation for the model by 26.49%,58.23%,and 36.98%,respectively,while ensuring the model recommendation accuracy.