Research On Deep Collaborative Recommendation Models For Sparse Data

Data sparsity has become an important challenge for recommender system currently.In order to alleviate the problem,academia has conducted deep research on recommendation algorithms,and the recommendation algorithms have improved and optimized from different perspectives by exploring various factors can influence the performance of recommender systems.Item correlation information can provide important information for recommender systems,however,most existing approaches exploit either global or local item correlations,rarely consider both global and local item correlations.In addition,the application of most traditional recommender systems is limited to a single domain,the datasets in a single domain tend to be very sparse.If we consider cross-domain recommendation,most existing cross-domain recommender algorithms are designed for the application scenarios where there are shared users or items between auxiliary domain and target domain,thus they don't apply to the application scenarios where there are no shared users or items between the two domains.If we transfer some information which is not relevant to target domain from auxiliary domain to target domain,this can result in negative transfer problem.Inspired by the above problems,we propose two collaborative deep recommendation model for data sparsity.One is collaborative deep recommendation with global and local item correlations.In order to alleviate data sparsity problems,we design a collaborative deep recommendation model by simultaneously considering both global and local item correlations.This model is also a general framework,we can use corresponding deep neural networks according to the type of item information.The proposed model combines rating information and item information by tightly coupled matrix factorization with denoising autoencoder,which makes the item latent features have an influence on the learning of item latent representations.In addition,this model provides a new perspective to integrate item correlation into recommendation model.We introduce manifold regularization to directly learn global and local item correlations from data.We conduct comprehensive experiments on three datasets at three different degrees of sparsity to confirm that GLICR can effectively alleviate data sparsity problem and is superior to existing state-of-the-art recommendation techniques.Another is deep selective transfer network for cross-domain recommendation.In order to select some user information which is important for target domain from auxiliary domain,we design a deep selective transfer network.In addition,in order to learn more effective item representation of target domain,we consider item content information of target domain.The DSTN model is formed by tightly integrating item content information into deep selective transfer network,which is a tightly coupled cross-domain recommendation method.This model is also a general framework,we can use corresponding deep neural networks according to the type of item information.Specifically,the proposed model uses a shared denoising autoencoder in auxiliary domain and target domain,and uses the rating matrices as the supervisory information by considering the learning process of autoencoder is unsupervised.By the deep selective transfer network,on the one hand,we can select the information which is useful for target domain from auxiliary domain,on the other hand,we can obtain user and item latent representations.Meantime,we extract item latent feature representation from item content information by a private denoising autoencoder for target domain,and we use the item latent feature representation to influence the item latent representation.We conduct comprehensive experiments on two public datasets confirm that DSTN is superior to conventional recommendation algorithms for a single domain and existing cross-domain recommendation algorithms,and DSTN is suitable for the application scenarios where there are no shared users or items between auxiliary domain and target domain.