Tag-aware Recommender System Based On Deep Learning

With the rapid development of the Internet,various information resources present blowout trend,making it difficult to provide users with quality information services relying solely on the search engines.To alleviate information overload,many companies have built recommender systems to provide users with accurate personalized recommendation.On some sites,users can freely annotate favorite online items with arbitrary social tags,which can not only indicate user preferences and opinions,but also summarize intrinsic features of items.Therefore,tag-aware recommender systems take the social tags generated by the collaborative behaviors as a kind of content information to achieve personalized item recommendation.To solve the problems of sparsity,synonymity and ambiguity brought by tags,some tagaware recommendation models based on deep learning were proposed,which improved the recommendation performance significantly.However,they also have some weaknesses that hinder their effect.They construct user representations by either explicit tagging behaviors or implicit interacted items,which is inadequate to capture multi-aspect user preferences.Besides,due to the multi-dimensionality of items,different users have diverse focuses on the same item for different reasons.Thirdly,it is difficult to alleviate synonymity and ambiguity relying solely on the full connection layer.To make full use of social tags to improve recommendation performance,we make an in-depth study on the weaknesses mentioned above of the existing studies and propose two novel tag-aware recommendation models which capture multi-aspect user preferences by attention mechanism and graph neural network,respectively.Moreover,the intersection of tags is also utilized to explore the deep reasons for user-item interactions.The main research results of this paper are as follows:(1)Analyzed the advantages and disadvantages of social tags and explored the weaknesses of the existing research.The strengths of deep learning in extracting latent representation and modeling feature interaction are fully developed and some advanced technology,such as attention mechanism and graph neural network,are utilized to design more effective tag-aware recommendation model.(2)Proposed attentive intersection recommendation model AIRec,which first leverages multi-layer perceptron to project the sparse tag vectors into a latent space,and uses a hierarchical attention network,which unifies both explicit tagging behaviors and implicit interacted items,for capturing multi-aspect user representations.Besides,the tags intersection is innovatively employed to enhance the learning of user-item conjunct features.Finally,factorization machine is leveraged to capture second-order interactive signals and achieve score prediction.(3)Proposed graph convolutional model TGCN,which uses the graph convolutional network to capture the high-order connectivity and takes into account users' tagging behavior and interactive items in the process of multi-hop neighbor information aggregation.More specific,proposed Trans Tag as embedding layer,leveraged “channels” to isolate different types of neighbors,designed an intra-channel weighted attention mechanism to identify the importance of neighbors and used multi-channel convolutional neural network to extract cross-channel feature,thus updating node embeddings.(4)Compared with the existing tag-aware recommendation model,the proposed models achieve better top-n recommendation effect on all evaluation metrics on two public datasets.In addition,a prototype system is designed to realize personalized item recommendation by using historical tagging records,verifying the good recommendation effect of our models.This paper first introduces the background and significance of tag-aware recommender systems,analyzes the current research status and deficiencies,then gives a brief overview of the tag-aware recommender systems,and puts forward the technical route of our research.Based on this,the attentive intersection model AIRec and graph convolutional model TGCN are proposed,then the concrete structure and training details of the models are discussed.A detailed experimental evaluation is carried out on two public datasets,a prototype system is designed and finally the future work is prospected.