Link Prediction And Recommendation Algorithm Based On Knowledge Graph

In the era of big data,the rapid development of Internet technology makes the public enjoy convenience,but also brings the problem of information overload.Recommendation system arises.By analyzing users ’historical behavior data,it excavates users’ potential needs,so as to help users to comprehensively and accurately screen content from massive data,and help users find the content they are interested in.However,the traditional recommendation algorithm has problems such as sparse data and cold start,which will make the recommendation effect not ideal and cannot bring users a comfortable experience.In order to solve such a problem,using knowledge graph as the auxiliary information of recommendation algorithm,the characteristics of rich semantic relationship,flexible semantic content,complex semantic correlation and other features in the data can be mined,so as to improve the accuracy of recommendation algorithm.However,most of the existing knowledge maps have problems with sparse data,incomplete information,and full noise.Knowledge graph mapping is an important area of current research that predicts missing relationships through known pairs of entities in the knowledge base.Since most knowledge graph completion methods only consider the direct relationship between relationship types and entities,ignoring the important influence of relationship semantics in the knowledge graph,it cannot fully mine the characteristics of triples.Aiming at these issues surrounding knowledge map completion and recommend two directions to discuss research,in order to solve the knowledge map information is not complete and map information utilization rate is low,the issues that led to the recommendation results inaccurate path link prediction model based on the Transformer and relationships,and to incorporate knowledge map completion technologies recommended tasks,An improved Ripple Net model is used to achieve accurate recommendation.The main contents are as follows:(1)Propose a link prediction model based on Transformer and semantic information.Existing models based on knowledge graph embedding deal with triples in the knowledge graph independently,ignoring the entity local neighborhood structure information,while the entity local structure in the knowledge graph may contain rich semantic relations and information.In this paper,a multi-path fusion strategy based on relationship context and relationship path is designed.The weight of each path is calculated based on attention mechanism,and the final representation of the relationship path is obtained by combining with the feature vector of the path.Relationship context information allows you to find the most important relationship path between entities.Experiments on different datasets have achieved good results.(2)Use Transformer model to capture path features comprehensively.Existing models generally use the representation of network learning paths such as RNN,LSTM or GRU to capture the similarity of different relational paths.However,when coding the path,RNN networks generally only consider the sequential dependence of adjacent entities in the path,and the forgetting gate,reset gate and other mechanisms will selectively discard the information that is not currently considered important in the process of learning the path.This can lead to information loss and the inability to adequately consider dependencies between non-adjacent entities.Therefore,Transformer encoder is used in this paper to encode the relational path,and the attention distribution among all entities in the path is calculated through the self-attention mechanism.When learning each relational path,the characteristics of different relational paths are captured,and a better path representation can be obtained,so as to better predict the relationship between entities in the triplet.Experiments show that learning path features through Transformer can effectively improve the prediction accuracy.(3)A recommendation algorithm model based on knowledge graph completion and item popularity is proposed.Firstly,the problem of data sparsity in the knowledge graph is alleviated to some extent by the knowledge graph completion technology.Then,in order to improve the ability of the recommendation system to mine unpopular items that users are interested in,Ripple Net model is improved by item popularity,which is introduced into similarity calculation and recommendation process as a weight factor to modify the score,so as to enhance the diversity of recommended items and achieve more accurate recommendation.The experimental results show that the recommendation effect is significantly improved in different application scenarios.