| With the rapid development of data science,recommendation systems have been proved to be the main technology to solve the data explosion.Among them,entity social relationship information is used as the main auxiliary information for current recommendation research.It has great advantages in many aspects such as mining of hidden information,improvement of recommendation models and recommendation scoring accuracy.In addition,the rapid development of graph neural network provides new methods and ideas for realizing many-to-many relationships among entities.It also brings more opportunities for the development of recommendation systems.In recent years,people have been gradually influenced by online socialization.Especially,the complex entity relationships between people can influence each other.Some entity relationships may negatively affect the recommendation results.It causes some invalid information to interfere with the effect of learning.It is easy to produce problems such as cold start and overfitting.In addition,the current entity social recommendation basically uses the long-term preferences of users in social relationships to model.There is a lack of integration of short-term preference information of items.Using long-term preference modeling alone does not allow for the complete construction of user preference models.In particular,short-term preferences are more representative for certain domains targeting immediate choice of items.In this thesis,the following two models are constructed with the above two problems in mind:(1)Research on social recommendation model based on enhanced neighbor perceptionIn this thesis,we construct a new social recommendation model,Research on social recommendation model based on enhanced neighbor perception(Sim-Graph Rec),to address the influence of invalid information mentioned above.The model judges the validity of neighbor relationships through real user-item interactions and friend trust relationships in the dataset.The effect of enhancing neighbor perception is achieved.In addition,in order to increase the recommendation accuracy.In this thesis,the closeness index and mapping are added to filter out the neighbors with real reliability.Then the attention mechanism is used to aggregate the enhanced social relationship features to achieve the effect of enhanced neighbor interaction perception information.Finally,this study establishes the project social space based on dyadicity and fuses it with the user social space,which can predict the ratings more accurately.In addition,this thesis conducts experiments on two publicly available datasets,Epinions and Ciao.Compared with the rest of the baseline models,the evaluation metrics of MAE and RMSE are improved by 0.81%-1.09% and 1.15%-1.41%,respectively,proving the effectiveness of the framework.(2)Research on long-and short-term social recommendation models based on convolutional gating and graph attentionThrough further research,it is found that in the fields of short video recommendation,news recommendation,etc.,the preferences of user entities often change with the changes of current events hotspots.Therefore,to address the problem of social recommendation in the above-mentioned fields,short-term preferences are not studied enough.In this thesis,we design a second model-a long and short-term social recommendation model based on convolutional gating and graph attention(CGA-Graph Rec).The temporal features of entities are extracted by constructing a convolutional gated attention network(CNN-GRU-ATT).Specifically,firstly,for the problem of accurate extraction of entity features,CNN networks are constructed to extract higher-level and more abstract features of itself and transform high-dimensional data into low-dimensional data;secondly,for the problem of social temporality,the use of gated units can effectively improve the operational efficiency;finally,for social relationship aggregation,graph attention networks are used to aggregate the social relationship information of entities.It constitutes the final feature representation of users and projects.In particular,when constructing the social information of project entities,modified cosine similarity is used to reduce the error caused by data insensitivity.Thus,prediction scores are derived.In this study,simulation experiments are conducted on two publicly available datasets(Epinions and Ciao).The results demonstrate the advantages of the proposed CGAGraph Rec model over other baseline models.It improves 1.06%-1.33 % and 1.19%-1.37%on two evaluation metrics,MAE and RMSE,respectively.The effectiveness of the model innovation was demonstrated. |
