Reserach On Recommendation System Based On Cross-modal Correlation Analysis And Adaptive Ranking Learning

The recommendation model is an effective information filtering mechanism,which finds valuable content from massive data according to user needs and recommends it to users in a various form.Its appearance meets the personalized needs of different users and brings great benefits to enterprises and society.However,the recommendation system is still in an immature stage,and there are still some problems,such as sparse data,low interpretability,underutilized cross-modal correlation between features and so on.Therefore,this paper carries out research based on the above existing problems.First,based on the original Movie Lens dataset,new multimodal datasets Movie Lens-100k-MPT and Movie Lens-1M-MPT are constructed by crawling movie text and poster information,which lays the foundation for later research work.Secondly,it is studied from the perspectives of explainable matrix,cross modal correlation analysis,adversarial learning and improved ranking learning function.The specific work of this paper are as follows:(1).Explainable bayesian personalized ranking recommendation model(EBPR): First,interaction table is constructed based on the interaction data between users and items in the Movie Lens dataset.By calculating the cosine similarity between items,the top N items with higher similarity are obtained.Then,the explainable matrix is generated according to the interaction matrix,and the explainable matrix represents the interaction probability between the user and the item neighborhood.Second,based on maximum likelihood estimation,explainable matrix is introduced on the basis of bayesian personalized ranking to increase the interaction between users and item neighborhoods,and realizing recommendation.The experimental results show that the explainable matrix can mine richer project information for users,and describe user preferences more accurately,which alleviates the cold start problem of adding new items to a certain extent.Compared with BPR model,the recommendation performance of EBPR model has been significantly improved,which shows that EBPR model has certain explainable.(2).Cross-modal correlation analysis adversarial bayesian personalized ranking recommendation model(CCABPR): Around the multi-modal Movie Lens-100k-MPT and Movie Lens-1M-MPT dataset,complementary text features are extracted based on bert model,and heterogeneous image features are extracted based on senet model;Based on gradKCCA and CCCA correlation models,the correlation between heterogeneous BERT text features and SENet image features are respectively mined.Adversarial learning strategy is designed to optimize the training process and better complete the recommendation task.The experimental results show that compared with the mainstream baselines such as NMF and VBPR,CCABPR model have significant advantages,indicating that text features,visual features,cross modal semantics and confrontation learning examples play an important role.Meanwhile,the CCCA model outperforms the gradKCCA model,indicating that the cross-modal semantics mined by the CCCA model is more discriminative.In addition,the model achieves better performance improvement on Movie Lens-1M-MPT dataset,which better alleviates the problem of data sparsity and improves the recommendation performance.(3).Cross-modal correlation analysis and adaptive ranking learning adversarial bayesian personalized ranking recommendation model(AABPR): Based on CCABPR model,AABPR model is proposed,and the pairwise ranking loss function in BPR model is improved.The model adaptively selects the ranking function point by point,pair by pair or any combination of the two according to the interaction relationship of project pairs,and finally completes the high-quality recommendation.The experimental results show that AABPR model is better than CCABPR,DMF,APR and other mainstream baselines,indicating that adaptive ranking learning is better than the point-by-point and pairwise ranking learning algorithms,and it plays an important role in recommendation.Meanwhile,the recommendation model has certain practicability.