Research On Recommendation Model Based On Deep Learning

The personalized recommendation technology mainly obtains the user's preference characteristics for items by recommendation algorithm and recommends items to users according to user's demands,preferences or other information.It plays an important role in alleviating the problem of "information overload".It has also applied to various fields and attracted much attention of researchers.In personalized recommendation system,the recommendation model based on collaborative filtering was used widely,especially matrix factorization(MF),SVD++ have made many achievements in Netfix competition,and pushed the research on collaborative filtering model based on matrix factorization to climax.However,the model just only considers the rating information between users and items,and the problem of data sparsely and cold start will greatly affect the recommendation performance of the model.In addition,as a shallow model,it can't to learn the deeper features of users and items,and the interaction information between users and items to make more accurate recommendation.With the development and breakthrough of deep learning technology,it has brought new opportunities and challenges for the research of recommendation model.The recommendation model based on deep learning utilizes the structure of multi-layer neural network,which can learn the interactive information between users and items non-linearly,obtain deeper and more abstract hidden feature representations,and show better result.But,most of models are based on the idea of matrix factorization,using a single rating information will reduce the recommendation performance of the model when encountering the problem of data sparsely;only the last latent representation of multilayer neural networks is interacted,without considering the feature representation learned by each network is also important.In model training,the same weight is applied to all feature interactions between users and items,but for different feature interactions,the importance is different.In this paper,research on the above problems,combines auxiliary information to alleviate the sparsely problem,constructs a new network structure to effectively utilize the latent representations learned by each layer of network,adding the attention mechanism to judge the importance of different feature interaction,and further improves the recommendation performance of the model by combining shallow model and deep model.The main work of this paper is as follows:(1)Considering the problem of the rating data sparsely and only using the last latent representation to recommend,we proposed a Multi-interactive Deep Matrix Factorization Model Based on Auxiliary Information.On the basis of deep learning recommendation model based on matrix factorization,we further utilize the rating data and fuse more auxiliary information(user/item attributes,comments,tags,etc.)as the model input,which contains some hidden preference features of users and items,it can not only alleviate the problem of data sparsely,but also extract more attributes of users and items from auxiliary information.Then,we construct parallel multi-layer non-linear network to learn latent representation of users and items respectively,and utilizing dot product operation for latent representation of users and items learned by each layer of network,considering different learning abilities of network layer to obtain interactive results of different layers.Finally,aggregate the interaction results of all layers as the model final result and predict the score.After compared the experiment on Movielens latest 100 K dataset and analyzed the model related parameters,the experimental results have proved that the proposed recommendation model can accurately predict the score.(2)Considering the importance of different interaction features,and further combine the shallow model and deep model to recommend,we proposed an attention multiinteractive neural matrix factorization model.Firstly,based on the multi-interactive network structure proposed by work(1),to obtain the interactive results of users and items learned by each layer of network;using the latent representations learned by multi-layer neural networks as input.Adding an attention network to learn an attention weight matrix,which is used to judge the importance of interaction between user and item features;utilizing the latent representations that learned from the multi-interactive network and the attention weight matrix,to obtain a result of deep model.Secondly,based on the idea of shallow matrix factorization model,we share the same embedding layer with the deep model mentioned above,obtain the latent factor representation of users and items,and directly use the dot product operation to obtain a result of shallow model.Finally,combine the linear shallow model and non-linear deep model,weighted the shallow result and the deep result to obtain the final results of the proposed model and make Top_N rank recommendation.After compared the experiment on Movielens 1m and Pinterest datasets,and analyzed the model related parameters.The experimental results have proved that the proposed recommendation model can more accurately to recommend items for users.