Research On Trust-aware Recommendation System Based On Deep Learning

In the field of recommendation system research,the collaborative filtering recommendation algorithm is the most widely applied and studied,and has achieved certain results in some recommendation tasks.However,the collaborative filtering recommendation algorithm only considers the rating data of users,which has the problems of data sparsity and cold start,so it is difficult to learn the strong correlation information between users and items.As a shallow recommendation model,the algorithm could not get the deeper hidden features between the user and the item,so the recommendation effect was not ideal.To alleviate the problems of data sparsity and cold start,trust information is added to the recommendation model as auxiliary information.However,the traditional trust-aware recommendation system uses the method of balancing user's own preferences and trusting user preferences for recommendation,and still belongs to the category of shallow recommendation models.The ability to mine deeper hidden features between users and projects is relatively weak.In addition,when the trust-based recommendation model deals with the similarity or attention distribution of users to trusted users,most of them are calculated by the user's common rating items,and the optimal attention distribution cannot be obtained to a great extent..Artificial intelligence technology represented by deep learning has brought an opportunity for the research of recommendation models.The deep learning-based recommendation model uses multi-layer neural networks to learn user and project interaction information,obtain deep hidden features,and obtain better recommendation results.Most of the existing models are based on the idea of matrix decomposition,and use a single rating data,which shows weak learning ability to the user's strong association information,and it is difficult to further improve the recommendation performance of the model.This thesis integrates trust information as auxiliary information,builds a new neural network model,combines trust information and rating information for recommendation,and alleviates problems such as cold start;introduces an attention mechanism to merge user short-term and long-term preferences,and further optimizes attention distribution to improve the performance of the recommendation model.The main work of this thesis is the following two aspects:?1?For most deep learning-based recommendation models relying only on user rating data,it is difficult to discover the strong association between users and items.Based on the user rating data,the thesis fuses the trust relationship between users,flexibly combines the preferences of users themselves and the preferences of users' trusted friends through the method of deep learning,and proposes a new trust-based neural collaborative filtering?TNCF?.Firstly,the generalized matrix factorization model is used to combine trust information and rating information to obtain recommendations based on trusted user preferences.Secondly,the non-linear kernel in the multi-layer perceptron model is used to learn the interactive features from the scoring data to obtain recommendations of the user's own preference.Finally,implicit predictions are made by aggregating the recommendations from the previous two steps.Trust information not only implies similar information between users,but also reflects certain social information,so it can enhance the ability to supplement interactive data,increase the probability of finding strong associations between users and items,and then improve the recommendation effect.This article chooses to perform comparative experiments on the FilmTrust,Epinions,and Ciao datasets.The TNCF model has significantly improved the Top-n recommendation effect compared to the baseline model.Compared with the suboptimal model,the HR @ 10 is the highest on the FilmTrust dataset.Increased 1.73%,NDCG @ 10 increased 2.88%;HR @ 10 increased 2.54%,NDCG @ 10 increased 1.62% on the Epinions dataset;HR @ 10 increased 0.94%,and NDCG @ 10 increased 1.43% on the Ciao dataset.?2?Aiming at the problem that previous trust-based recommendation algorithms cannot optimize attention distribution,the thesis proposes using the attention mechanism to calculate user trust,that is,the attention of trusted users to capture the influence of friends' preferences.At the same time,in order to further study the changes in user taste,the thesis introduces time sensitivity and adopts the self-attention mechanism to capture the short-term preference changes of users.Based on the above two points,the thesis proposes an attention-based trust-aware sequence recommendation model?ATRec?.First,learn the short-term user preferences based on the user interaction sequence by setting the query,key,and value of the Self-Attention block as the user's recent L interaction sequences.Then,through the attention mechanism Allocate users 'attention to different trusted friends,learn users' long-term user preferences based on trusted friends.Finally,parameter ? is used to balance the short and long term user preferences to make recommendations.The experimental results show that the top-n recommendation effect of the ATRec model is also significantly improved compared to the baseline model.On the Flixster₁ dataset,HR @ 10 is increased by 1.82% to 11.47%,and NDCG @ 10 is increased by 1.47% to 23.33%.In the Flixster₂ dataset HR @ 10 increased by 3.07% 4.04%,NDCG @ 10 increased by 1.33% 8.75%,and training efficiency has also been improved.