Research On Cross-domain Recommendation Algorithm Based On Knowledge Transfer And Aggregation

In the Internet era of information overload,how users obtain useful information from massive information has become a hot research issue.Recommendation systems can alleviate the problem of information overload to a certain extent,but the performance of traditional recommendation algorithms is limited by the sparseness of users and projects in a single domain Sex is also limited by the cold start problem faced by new users without historical behavior.Due to the similarity and relevance of user group preferences and item attribute categories in different fields,cross-domain recommendation can predict the user behavior in the target field by learning the knowledge of the auxiliary field,thereby enriching the sparse data in the target field to improve the accuracy of recommendation.Based on this,this paper carried out cross-domain recommendation algorithm research on knowledge transfer and aggregation,the main work is as follows:(1)In order to solve the problems of user cold start,the data resulting from feature extraction after inter-domain scoring matrix is directly spliced,and it is not easy to generalize to new domains,etc.In the scenario where users partially overlap,this paper proposes from the perspective of group effects A cross-domain recommendation algorithm ATCF(Aggregation and Transfer Collaborative Filtering for cross-domain recommendation)is proposed.Different from the existing cross-domain recommendation algorithm,in the representation and learning of common knowledge and individual knowledge,the knowledge of the auxiliary domain and the target domain is fully fused,through two-stage stitching and two Filling and getting a common knowledge representation alleviates the problems of sparse data and cold start of users.Through knowledge transfer,the personalized knowledge representation of differentiated overlapping users and non-overlapping users is constructed,which effectively avoids negative transfer.Through the common knowledge representation based on knowledge aggregation and the personal knowledge representation based on knowledge transfer,the recommendation performance is improved.(2)Aiming at the negative migration caused by ignoring the interpersonal user personality characteristics and inter-domain user interaction features with the project,and the problem of user cold start,modeling for rich users,extracting the individual characteristics of each user,overlapping in the user part In this scenario,this paper further proposes an attention-feature transfer based on mapping for cross-domain recommendation(AFTM)from the perspective of individual effects.The algorithm models users through matrix decomposition and attention mechanism,integrates user-item interactions to construct user feature vectors,and then constructs feature maps between domains through neural networks to capture non-linear interactions between user features.The integration of user interaction characteristics between domains alleviates the problems of data sparseness and user cold start,and to a certain extent avoids negative migration,improving the accuracy of recommendations in the target domain.This paper has conducted experiments on the standard dataset Movielens and the real dataset Douban books and Douban movies.The experimental results show that the two cross-domain recommendation algorithms proposed in this paper are more accurate in recommendation than single-domain recommendation algorithms and other cross-domain recommendation algorithms.There are obvious advantages in the performance of the cold start and cold start users,which can effectively improve the recommended performance in the target domain.