Reserach On Recommendation System Based On Cross-modal Semantic Mining And Generative Adversarial Networks

With the advent of the era of big data,to actively deal with the problem of massive and unstructured "information overload",the application of recommendation system in various Internet products has become increasingly widespread,and they have brought important influences to people's daily lives.The research work of recommendation system has been developed for more than ten years,but there are still some problems such as data sparseness,insufficient consideration of users' potential preferences,and insufficient mining of the semantic correlation among heterogeneous image features.To address the above problems,on the basis of the external information sources(images)and the MovieLens dataset,a new multi-modal dataset called MovieLens-MP(MP refers to Movie Poster)is built first to build a strong data foundation for further recommendation.Then,we start our research work on recommendation system from the perspective of hard negative sampling,adversarial learning,modified VBPR model,image feature learning,and cross-modal semantic mining etc.Finally,the proposed models are trained based on the generative adversarial network(GAN).The main research sub-tasks are presented as follows:(1).A novel recommendation system based on the GraphGAN model.Based on the original MovieLens dataset,a novel recommendation system is proposed from the perspective of graph model: a bipartite graph of recommendation data is constructed first.Then the GraphGAN-based recommendation system is created according to the similarity probability of the nodes and edges in the bipartite graph.Experimental results demonstrate that the GraphGAN model can make full use of the structural information between nodes to boost the final performance.The proposed recommendation system is trained based on the generative adversarial network.Compared with the traditional model,recommendation performance is improved to a certain degree.(2).A novel recommendation system based on the UPM-GAN(UPM means User Preference Mining)model: Based on the original MovieLens dataset,user's potential preference is mined out from two perspectives: the state-of-the-art triplet loss algorithm is used to complete hard negative sample mining.The well-known SVD ++ algorithm is used to create the generation model of the UPM-GAN.The UPM-GAN model is trained under the GAN framework to complete better recommendation.Experimental results demonstrate that those mined negative samples can establish positive samples by a kind of inverse incentives to better characterize users' preferences;The SVD++ algorithm contains various implicit parameters and bias information,which help more accurately describe users' preferences;The proposed UPM-GAN based recommendation system achieves excellent performance and its convergence speed is faster.Moreover,the training process of the recommendation system is smoother.(3).A novel recommendation system based on the VABPR(Visual Adversarial Bayesian Personalized Ranking)model: To further alleviate the data sparseness problem,on the basis of the new MovieLens-MP dataset,the efficient matching kernel algorithm is first used to complete image feature learning;then the state-of-the-art VBPR model is combined with a novel adversarial learning strategy,which means visual features are absorbed into the recommendation framework and the adversarial learning is completed more robustly;Finally,the VABPR-based recommendation system is trained.Experimental demonstrate that compared with other models including UPM-GAN and GraphGAN,the final recommendation performance is further improved.The proposed visual features and adversarial learning strategy play very important roles in recommendation.In addition,the sparser dataset(MovieLens-1M-MP)can obtain better performance,which means the data sparseness problem is suppressed to a certain degree.(4).A novel recommendation system based on the MVABPR(Multi-model Visual Adversarial Bayesian Personalized Ranking)model: To make full use of the semantic correlation between heterogeneous image features,the MVABPR model is proposed based on the above-mentioned VABPR model and cross-modal correlation analysis.The canonical correlation between heterogeneous image features is mined out and cross-modal semantic is generated in turn to better characterize images' content.This indicates the visual representation of the proposed VABPR model is updated and high-quality recommendations can be obtained.Experimental results demonstrate that the proposed MVABPR model is superior to several mainstream baselines such as UPM-GAN,ABPR,and VABPR.More importantly,the new model can obtain better performance on the sparser dataset(MovieLens-1M-MP),the MVABPR-based recommendation system has larger practical value.