Research On Collaborative Filtering Recommendation Algorithm Based On Deep Neural Network

With the rapid popularity of the Internet,users can browse an increasingly rich variety of information,which brings enough convenience for their choices.At the same time,users also suffer from the difficulty of getting the information they want in the face of different kinds of information,thus failing to meet their practical needs.The emergence of recommendation system has solved these problems by providing personalized services for users to find useful information for themselves,thus improving their experience and making a big change in their lives.Currently,collaborative filtering techniques have been widely used in recommendation systems.Due to the many drawbacks of traditional recommendation algorithms,they cannot achieve satisfactory results in the recommendation field.On the one hand,deep neural networks can provide in-depth representations of user features and item features based on their nonlinear characteristics;on the other hand,deep neural networks can automatically learn feature representations from user-item interaction data,map different feature information to the same hidden space,and then obtain representations of feature information.Therefore,this paper combines deep neural networks for recommendation based on traditional recommendation algorithms to make full use of the user-item interaction data.The main work of this paper is as follows:(1)A new neural collaborative filtering recommendation algorithm based on channel attention(NCFCA)is proposed to address the problems of how to improve the relevance of embedding dimensions between users and items,the generalization ability of the model and how to accurately model users’ preferences for items in implicit feedback.Firstly,the feedforward attention mechanism is used to assign personalized weights to different interactive items during the interaction between users and items,so as to influence the degree of user preference to the items and obtain the characteristic information of users and items accurately and efficiently.Secondly,the convolutional neural network is used to enhance the relevance between users and items,and the channel attention mechanism is added to the convolutional neural network to mine rich semantic information.Finally,the generalized matrix decomposition method is used to alleviate the data sparsity problem arising from user-item interaction and to fuse three different modules(A-MLP,E-CNN,GMF)together to achieve top-N recommendation.(2)To address the problems of user’s ability to express different items at item level and to distinguish the importance of different historical items in implicit feedback to construct user preferences,this thesis proposes a neural collaborative filtering recommendation algorithm based on autoencoder and attention factor(NCFAAF),which uses a deep autoencoder to extract the initial representations of users and items to obtain deeper feature information,effectively reducing the impact of redundant information on recommendation performance.In terms of items,by introducing attention factor is introduced to narrow down the difference of users on historical item sets and improve the judgment of user preferences.In order to have a fuller understanding of the collaborative filtering technology,this thesis proposes the neural collaborative filtering recommendation algorithm based on channel attention and the neural collaborative filtering recommendation algorithm based on autoencoder and attention factor,and conduct a large number of relevant experiments to verify these two models on real data sets.The experimental results are analyzed to verify that the proposed models can improve the shortcomings of the traditional recommendation algorithms to a certain extent,and further demonstrate the effectiveness and rationality of the proposed models.