Research On Path-based Recommendation Technology With Knowledge Graphs

The recommendation algorithm can accurately locate user's preference and target products in Internet Applications.As a graph structured human knowledge database,knowledge graph(KG)organized in triples is well-structured and has rich semantics,which is widely used in recommendation such as e-commerce networks,music players,etc..By obtaining paths between users and products in KG,the explainability of the recommendation algorithm has been greatly enhanced.However,in real-world scenarios,KGs are often extremely large in scale and complex in structure.It is difficult to quickly extract and effectively represent exponential-level paths,which reduces the credibility of recommendation algorithms.How to implement a fast,accurate,and explainable recommendation algorithm in a highly complex and large KG has become a research focus.Path-enhanced Recurrent Network(Pe RN)greatly reduces the cost of obtaining paths from the KG and effectively improves the model's accuracy and explainability.Specifically,Pe RN adopts an efficient path extraction strategy,which can simultaneously perform depth-first search and matching from entities at both ends with the aid of meta-paths,making it possible to quickly obtain paths from a large-scale KG.In addition,Pe RN uses a recurrent network encoder based on bidirectional Long Short-Term Memory networks and an entropy encoder based on information gain to encode the path and its meta-paths,which implements an end-to-end model and guarantees the accuracy and explainability of the recommendation algorithm.To validate the effectiveness and efficiency of the algorithm,two real-world KG recommendation datasets,named KKBox music dataset and IM-1M movie dataset,are adopted in the experimental part.The results show that compared with the existing methods,Pe RN is more than two orders of magnitude faster,the accuracy is increased by5% to 10%,and the recommendation results are more explainable.