Research On Collaborative Evolution Recommendation Based On Self-attention

In order to solve the problem of sparse data in traditional recommendation algorithms,the fusion of the relationship information between temporal sequence information and knowledge graph has become an important way to solve this problem.The method of fused temporal information can model the evolution of user interests,but ignores the problem of semantic relationship between item interaction sequence and information loss caused by parameter sharing.The method based on knowledge graph is used to mine semantic paths and topological relationships,ignoring temporal information and existing semantic paths on knowledge graph construction.Aiming at the above problems,this paper proposes a self-attention based collaborative evolution recommendation algorithm,which alleviates data sparseness and improves recommendation performance.The recommendation algorithms which fuse temporal information,cannot effectively obtain and utilize the semantic relationship,but also bring information loss caused by parameter sharing.Thus,a collaborative evolution recommendation algorithm based on semantic relationship extraction is proposed.It firstly uses sequence features and semantic relationship features extracted by the recurrent neural network and self-attention mechanism as short-term features of users and items for solving the problem of incomplete semantic extraction of interaction sequences and information loss;secondly,it uses matrix factorization to extract long-term features of users and items;and then fuses long-term and short-term features of users and items and uses multi-layer perceptron to make score prediction for solve the ranking loss caused by inner product.The evaluation results of the algorithm on public dataset show that the algorithm can effectively improve the accuracy of recommendation.Previous recommendation algorithms based on knowledge graph ignore the temporality and the influence of the existing semantic path information between entities.Therefore,an temporally incremental construction algorithm based on semantic path extraction for knowledge graph was proposed.For each user item interaction event,firstly using multi-layer perceptron to extract the effect of the interaction itself on node vector updates;secondly,use recurrent neural networks to extract semantic information of semantic paths with different lengths,and use attention mechanisms to distinguish the effects of different length paths;Finally,they are merged into the update process of the node representation.The experimental results in two public datasets show that it can effectively improve the performance of recommendation.