The Research On Recommendation System Based On Flow And Generative Network

Collaborative filtering(CF)is the most widely used and successful technique in personalized recommendation systems.However,existing collaborative filtering based methods still have certain limitations.In recent years,deep generative models such as,Variational Autoencoders(VAEs)and generate adversarial networks(GANs),have been introduced to CF-based recommender systems,producing models with promising results while improving ranking performance and model robustness.However,these models are inherently deterministic and cannot be generalized to estimate the recommendation uncertainty.In addiction,these approaches typically lack characterization of explicit density,making it difficult to directly model user-item interactions.This thesis addresses aforementioned issues through modeling and estimating implicit feedbacks of user-item interactions.Towards this goal,we extends the VAE with the capability of stochastic and amortized inference,enabling better variational approximation and better recommendation performance.Specifically,we made following contributions.For the item recommendation service in the typical on-line applications,this thesis presents collaborative autoregressive flows(CAF),a novel collaborative filtering-like model that leverages the Bayesian inference and autoregressive flows for item recommendation.CAF is a non-linear probabilistic method that can provide an accurate representation of uncertainty representation and latent variable inference in item recommendations.Compared with the prior approximation of agnostic presumption used in existing depthgenerating recommendation methods,CAF is more effective in estimating the posterior of the probability,and can improve and interpret the representation learning of latent factors with autoregressive flows.The proposed model allows flexible and tractable probabilistic density estimation by exploiting the flows to approximate the true posterior of stochastic latent factors,largely alleviating the inference bias in existing Bayesian recommendation models and improving the recommendation performance.Hybridizing two autoregressive flows endows CAF with the benefits of both components,i.e.,the efficiency on variational inference and sampling of data.For the point-of-interest(POI)recommendation service in Location Based Social Networks(LBSN),this thesis presents a novel CF model(WaPOIR)that leverages Bayesian inference and probabilistic generative model with VAE-based networks for POI recommendation,which can capture the non-linear user-POI relationships.Moreover,WaPOIR can alleviate the data sparsity and cold start problem by incorporating the geographical influence along with user preference and social influence.In addition,our novel neural attention network can simultaneously capture users' general and current interests.Different from conventional VAE-based recommender systems,WaPOIR learns the Gaussian distribution embedding of input data in the Wasserstein space.WaPOIR satisfies the triangle inequality which can well preserve the transitivity in LBSNs.More importantly,it introduces the representation uncertainty of both users and POIs,as well as their similarities which,in turn,improves the learning of their latent representations and interactions.