Research On Social Recommendation Algorithm Based On Disentangle Representation Learning

Since the 21st century has fully entered the information era,the existence of recommendation systems has effectively solved the information overload problem in the face of the growing mass data,however,recommendation algorithms based only on useritem rating matrices have data sparsity and cold-start problems,the introduction of social relationships provides an effective way to solve these problems,and the social recommendation problem has therefore received extensive attentions from researchers.There are often multiple types of user social relationships hidden in social networks,and directly integrating social relationships into the scoring matrix may lead to degraded recommendation performance.Extracting user relationships related to ratings from social networks is the key to improve the recommendation performance.In this paper,we use the feature of disentangled representation learning to learn disentangled representations with separated semantics and apply it to the social recommendation algorithms.And using the disentangled social representations related to recommendation tasks in social networks to improve recommendation performance.The main work of this paper is as follows.(1)A social recommendation model with disentangled recommendation-related social information(DSSR)is proposed to address the problem that current social recommendation models based on graph neural network fuse recommendation-unrelated social representations when modeling user interests and affect the recommendation performance.The model first learns recommendation-related disentangled representations of users in social space based on disentangled graph convolutional networks,then fuses them with user interest representations in interactions space with items,and finally introduces a social network reconstruction task to improve the disentanglement quality of user representations in social space.Experiments on two publicly available datasets show that the model proposed in this paper achieves better recommendation performance than other methods in both general recommendation scenarios and user cold-start recommendation scenarios.(2)The DSSR model not only needs to optimize the parameters of social network disentangled layers and interaction network convolution layers at the same time,but also cannot learn user higher-order relationships in layers,for which a dual-domain fusion social recommendation algorithm based on disentangled representation learning(DSSR-DF)is proposed.The algorithm first uses dual-domain shared vectors to fuse user interests in both social space and interaction space,then achieves recommendation-related user interest extraction by disentangling and reconstructing social networks,and finally achieves higherorder relationship fusion by multi-layer network stacking of social network disentanglement and interaction graph convolution.Experiments on sparse and dense recommendation datasets show that DSSR-DF can effectively reduce the hyperparameters of the model without losing recommendation performance compared with DSSR.(3)We designed and implemented a software for recommending tourist attractions based on users’ social relationships.The software uses users’ social relationships and their interaction records with attractions to reason about users’ interests,and provides users with personalized recommendation services for tourist attractions based on social recommendation algorithms with collaborative filtering strategy.