Research On Top-N Recommendation Algorithms Based On Side Information And Variational Auto-Encoders

Personalized recommendation has continuously received attention due to its great commercial value in business.As the number of commodities grows rapidly in recent years,Top-N recommendation is becoming a new research focus because of its accurate push characteristics.Deep neural networks can automatically excavate the behavior patterns from the historical interaction records,which has achieved excellent results in related tasks.Among them,the variational auto-encoders are superior for learning to rank and recommendation on massive data.However,the existing Top-N recommendation methods still have limitations on the efficient integration and utilization of side information,which further affects the quality of recommendation services.To solve the problem,this thesis studies the effective fusion of side information based on the variational auto-encoders to learn a more meaningful latent factor from user-item interactions.The main work of this thesis is as follows:1)Targeting at the problem that splicing and stacking network structure machinery can not make full use of side information,this thesis proposes a Top-N recommendation model of variational auto-encoders with attribute features.This method encodes user attributes into condition vectors,and integrates into loss function as a label verification signal to cluster the latent factor representation.In addition,this method utilizes the characteristics of the generative model by separating attributes to learn,and integrate the results of multiple prediction pools to propose improvements.Experimental results show that the method has a good modeling ability of the preferences among the user groups when compared with related methods,and favorable learning performance.2)Targeting at the problem that association degree between user behavior and complex auxiliary information is not the same,this thesis proposes a Top-N recommendation model of variational auto-encoders with attenion mechanism.This method captures the relationship between user's representation and side information through a sub network,balancing the fusion weight of attributes in the main network.In addition,this method tries to construct combination of features in the high-dimensional embedding space,helping mining the promotion of side information at a finer scale.Experimental results show that the method has a better interpretability of feature representation correlations when compared with related methods,and enhanced recommendation quality.3)Targeting at the problem that implicit feedbacks can not fully reflect the user's preference on items,this thesis proposes a Top-N recommendation model of variational auto-encoders with explicit feedbacks.This method integrates two kinds of feedback information into the network by reconstructing explicit score and implicit click records at the same time.In addition,this method prioritizes the most likely items to browse and reorders them according to the predicted scores,followed by proposing an adaptive algorithmn at the re-ranking stage of recommendation.Experimental results show that the method can collect the feedback information effectively and significantly improve the final recommendation performance when compared with related methods.