Research On Collaborative Filtering Recommendation Algorithm Based On Attention Mechanism And Heterogeneous Network Evolutionary Clustering

The rapid growth of e-commerce platforms has provided users with unprecedented services and products,and a major bottleneck of these platforms is that users need to deal with a large amount of redundant information before making decisions.At this point,the advent of the recommendation system provides users with the appropriate items,and make the ideal decision.The recommendation system fully mines the user's interests and preferences through the user's historical data,and then recommends items for the user according to these preferences.However,with the rapid development of network technology,the number of users and items grow explosively,which brings so many problems such as cold start and data sparsity to the recommendation system.To this end,researchers carried out a series of studies.Among the numerous recommendation algorithms,collaborative filtering recommendation algorithm based on clustering is very popular because it takes into account the user's community relationship and can reduce the recommendation time.However,the existing algorithms based on evolutionary clustering only rely on single-layer network,that is to say,they are limited by a single node.And it doesn't integrate the user's attention into the recommendation system.In addition,most of the user similarity indicators are limited to describe the symmetric relationship,and ignore the user preferences.To address the above problems,this thesis proposes an evolutionary clustering recommendation algorithm based on heterogeneous network and attention mechanism.The specific research contents are as follows:(1)Personalized recommendation algorithm in double-layer network based on vector dynamic evolution clustering and attention mechanism.Firstly,in the Chapter 3,this thesis constructs a heterogeneous double-layer network by using the users attribute information,items attribute information and rating information,and introduces the forgetting function.Secondly,the previous evolutionary clustering is only based on a single-layer network,and the node state value is a scalar,which cannot fully represent the change value of the node states.Therefore,a vector dynamic evolutionary clustering method is proposed to cluster users and items respectively,so as to obtain the appropriate user community and item community.Then,the similarity matrix within each user community is calculated,and then the prediction score of target users to target items is obtained by the proposed double-layer network rating prediction method.Finally,items with high scores are recommended to target users by predicting their scores.By means of Lyapunov theory,it is proved that the proposed vector dynamic evolutionary clustering equation can converge after a period of oscillation.The experimental results of comparison with the existing better algorithms on three datasets show that the algorithm has certain advantages in improving the recommendation performance,can solve the cold start problem,and alleviate the data sparsity problem to a certain extent.(2)Hypergraph-based algorithm for personalized recommendation via game evolutionary clustering and attention mechanism.Because the attention mechanism and similarity measure in the first research method do not take into account the user's preference,and the state value of evolutionary clustering is linear with the state value of the last evolution,it does not conform to the actual situation and other issues.In the Chapter 4 of this thesis,the attention mechanism based on user rating matrix is proposed to obtain the attention matrix more in line with the actual situation with the help of heterogeneous hypergraph network model.Then,a new evolutionary clustering method is proposed by combining the evolutionary clustering equation and the payoff function in game theory,and its stability is proved.By means of payoff function,the user's earnings obtained from its own effects are no longer linear,which makes it closer to the evolution of the hypergraph model.Then,the prediction score is obtained by the results of hyperedge clustering and the improved similarity calculation method,so as to carry out the user-based collaborative filtering algorithm.Finally,compared with the better clustering algorithm and recommendation algorithm on four datasets of different scales and types,the experimental results show that the proposed algorithm has great advantages in predicting scores and recommending items.(3)Attention-based adaptive evolutionary clustering in heterogeneous hypergraph model for personalized recommendation algorithm.In view of the traditional graph network nodes only exist in pairs,this thesis establishes a heterogeneous hypergraph model which is the same as that in Chapter 4 based on the user's historical evaluation information.In addition to expressing pairs of nodes,this model also includes a relationship of more than two nodes,which makes the algorithm mine more complex topological relationships.Secondly,a new attention mechanism is proposed by using the attribute information of the item and the score matrix of the user,and an asymmetric attention matrix is obtained.Thirdly,an adaptive evolutionary clustering equation is proposed,and attention matrix is used as its clustering principle for hyperedge clustering.Then,a similarity measurement strategy based on user preferences is proposed,and the similarity matrix is calculated in each hyperedge cluster.Finally,according to the similarity between the target user's hyperedge and the nearest neighbor,the user's rating on the item in the testset is predicted.The effectiveness of the proposed algorithm is verified by experiments with the better algorithm.From the relevant experiments and results of the above three research contents,the three methods in this thesis are of great help to address the problems of cold start,data sparseness,and real-time,and are of great significance for reducing prediction errors and improving recommendation accuracy.