Clustering Based Collaborative Filtering And Graph Self Encoder Research On Personalized Recommendation System

With the rapid development of machine learning and artificial intelligence technologies,the explosion of information on the Internet has led to problems such as high volume,poor quality,and low value of information.In order to better access high-quality information,recommendation systems have emerged,which take users’ historical behaviors,interests,and demographic information as inputs and employ these data to train models and generate recommendation lists that maximize user interest and satisfaction.In this thesis,two perspectives of traditional recommendation algorithms and deep learning recommendation algorithms are investigated,and the following two parts are discussed:In the first part,a recommendation system based on improved sparrow search with clustered collaborative filtering is investigated.In the field of traditional machine learning recommendation algorithm,this thesis firstly adopts the combination of K-means clustering algorithm and collaborative filtering algorithm to solve the problem of data sparsity.Then,a new swarm intelligence optimization strategy,the sparrow search algorithm,is introduced for problems such as K-means clustering algorithm centroid finding,and the algorithm is improved by adding Chebyshev chaos initialization population,inertia weights and adaptive selection variation strategy.The experimental results show that compared with the K-means-based collaborative filtering algorithm,the recommendation algorithm based on the improved sparrow search and clustering collaborative filtering proposed in this thesis effectively alleviates the problem of falling into local optima due to the overly concentrated selection of initial clustering centroids.At the same time,the method can also better determine the number of clusters and has a large improvement in the accuracy of recommendations.In the second part,the probability-based self-encoder model of random wandering graph is investigated.In the field of deep learning recommendation algorithms,this thesis proposes a probability-based random wandering graph self-encoder model for the problem that existing graph self-encoders cannot capture the information correlation between nodes in the graph data structure well.This model firstly improves the random wandering algorithm by using the number of neighbors of a node to assign a weight to the current node,which in turn optimizes the transfer probability.Secondly,in order to better combine the graph convolutional network and the random wandering network,an adaptive combination method is proposed to optimize the weights through the loss function.To demonstrate the effectiveness of the method,the final representation of the graph network is applied to the link prediction task in this thesis.Experimental results show that the method proposed in this thesis achieves superior recommendation results compared with the original graph self-encoder.Among them,the AUCs on Cora,Citeseer and Pub Med datasets are improved by 3.07%,6.7% and 12.6%,respectively,and the APs are improved by 3.15%,6.45% and 1.13%,respectively.