Research On Deep Recommendation System Based On Knowledge Graph

The knowledge graph structure has strong correlation with the user-item interaction matrix in the recommendation data.So fusing the knowledge graph into the recommendation system can introduce stronger semantic relationships to the recommendation data.And at the same time,the hidden features of users or items can be obtained through the topologies of the knowledge graph to improve the accuracy,diversity and interpretability of the recommendation results.There are three strategies for fusing knowledge graphs into recommendation systems,namely vectorization-based fusion,path-based fusion and joint fusion.The traditional vectorization-based fusion method reduces the dimensionality of users or items to fit into the framework of the recommendation system.The traditional vectorization-based fusion method reduces the dimension of users or items in order to integrate the low-dimensional vectorization results into the framework of recommendation system.But this method ignores the connectivity of the knowledge graph and the characteristics of the knowledge graph information itself.The traditional path-based fusion method converts user-item interaction data into a bipartite graph and predicts recommendation results by calculating the similarity of the paths in the graph,which provides interpretability of recommendation results,but the model is usually not universal due to the difficulty of setting up the meta-paths.The joint fusion method is a combination of the above two methods and also a hot research topic in recent years.At present,there are some shortcomings in the joint fusion method,which are reflected in many aspects,such as learning difficulties of multiple relationships in the process of knowledge graph vectorization,lack of universality in knowledge graph meta-path design,lack of knowledge graph data meaning,knowledge graph information loss caused by low-dimensional graphping in the process of knowledge graph vectorization and recommendation system fusion,etc.In this paper,we analyze the joint fusion method of knowledge graph and recommendation system,optimize and design the model from the perspectives of optimizing the complex relationship processing process of knowledge graph,extracting the meaning of knowledge graph nodes and fusing the knowledge graph with recommendation model in multiple ways,aiming at solving the shortcomings of the current joint fusion method and extracting the topological structure information and semantic information of the knowledge graph constructed by recommendation data as comprehensively as possible and incorporating them into the recommendation system.In order to solve the shortcomings of the current joint fusion method,the topological structure information and semantic information of the knowledge graph constructed by the recommendation data are mined as comprehensively as possible,and integrated into the depth recommendation model to improve the basic performance of the recommendation model.This paper examines the following specific contents:1.A Depth Recommendation Model for Knowledge Graph Convolutional Graph Networks Based on Wandering StrategyIn the process of vectorization-based knowledge graph eigenvector and recommendation system fusion method,the structural information of the knowledge graph itself is often ignored.In the process of path-based fusion method,although it is difficult to set the path,there are also problems such as difficult path setting and poor universality because the path is difficult to set.In this paper,we propose a depth recommendation model of knowledge graph convolutional graph network based on wandering strategy,KGCNpro model,which belongs to the joint fusion method of the two fusions.Firstly,in order to better the universality of knowledge graph migration,the knowledge graph composed of recommendation data is reconstructed from a heterogeneous network structure to a homogeneous network structure.And the model is improved under the wandering strategy according to the principle of "entity-relationship-entity" in order to better fit the model of the knowledge graph triad.By reconstructing the knowledge graph,the distribution of training weights of different entity relationship attributes is solved.Aiming at the network structure of the knowledge graph,a targeted walking strategy is designed to solve the training problem of complex relationships of the knowledge graph.Two training models and optimization strategies are proposed to improve the accuracy of feature extraction of the knowledge graph.According to the result of the reconstructed knowledge graph walk,the entity nodes are reconstructed,and the walk sequence is directly used as the adjacency matrix of graph convolution neural network.Considering the breadth,the topological structure information of the knowledge graph can be retained to the greatest extent,and the walk strategy can also consider the breadth and depth of the graph structure of the entity nodes,and the knowledge graph is also depthly mined.The.KGCNPro model combines graph-based embedded representation and dissemination ideas,which enriches the diversification of integration modes in the process of knowledge graph integration and in-depth recommendation,and improves the accuracy of feature extraction of knowledge graph and the accuracy of recommendation results.2.A Multi-Modal and Multi-Task Depth Recommendation Model Based on Knowledge Graphs as Auxiliary InformationIn the model that combines knowledge graph with depth recommendation,items and subattributes of items are usually used as data to build knowledge graph.Although knowledge graph as auxiliary information can improve the recommendation accuracy,some information is not suitable to be expressed in the form of knowledge graph,that is,the node meaning of knowledge graph itself.In this paper,a multi-modal and multi-task depth recommendation model based on knowledge graph as auxiliary information,SI-MKR model,is proposed.In this model,the attributes of knowledge graph items composed of recommended data are divided into text type attributes,multi-valued attributes and other types of attributes by classifying and discussing the attribute types of items by means of multi-modal feature learning.Text usage adopts Text-CNN model for depth semantic mining.Multi-value type attributes are indexed in sequence,initialized with one-hot coding and vector addition,which together with the vector results of text type attribute training are used as the input of multi-layer perceptron layer,and then the output results of multi-layer perceptron layer are cross-trained with the knowledge graph structure composed of other types of attributes,so as to integrate the multi-modal representation learning results with the topology structure of knowledge graph through multitasking and improve the accuracy of depth recommendation.Experiments show that compared with other depth models in recommendation performance,the SI-MKR model proposed in this paper has a greater ability to combine multi-modal thinking with multi-task thinking by combining sequential training and alternate training,which improves the learning ability of one-to-many and many-to-many relationships in the process of knowledge graph vectorization,and improves the mining ability of data's own characteristics while retaining the topological structure of knowledge graph.3.Depth Recommendation Model Based on Ripple Propagation and Multi-Task Learning of Knowledge GraphIn the joint fusion method of knowledge graph and recommendation system,historical interaction data between users and projects or neighborhood data of project data can be used as auxiliary information in the recommendation process.At present,the fusion mode of joint fusion method is relatively simple,which leads to incomplete mining of potential information in knowledge graph.How to integrate knowledge graph with recommendation system to the greatest extent is worth considering and studying.To solve this problem,this paper proposes a depth recommendation model,Ripp-MKR model,which combines ripple propagation with multi-task learning based on knowledge graph.This is a depth recommendation model combining joint learning and alternate learning.In this model,firstly,the knowledge graph is cross-trained with the recommendation data,and the feature representation of the knowledge graph is integrated into the recommendation by multi-tasking.Secondly,the historical interaction information between users and projects is used to obtain the user's context and then the user's feature representation through the idea of ripple propagation.In Ripp-MKR model,the role of knowledge graph is maximized,the head of knowledge graph and project data are alternately trained,and the tail of knowledge graph is involved in the formation of user's feature vector,which combines joint training and alternate training to improve the performance of recommendation system.The innovation points of the combination model based on knowledge graph and depth recommendation of the three joint fusion methods proposed in this paper are summarized as follows: all belong to joint fusion methods.The problem of dealing with the complex relationship of knowledge graph in the process of knowledge graph vectorization,the problem of missing the semantics of knowledge graph nodes themselves,and the loss of information in the process of knowledge graph vectorization were solved.The problems that exist in the process of combining knowledge graph with recommendation system,such as the inadequate combination of topological information representing learning results with knowledge graph,incomplete information dissemination,and single training method,are solved.A large number of experiments have been carried out on the real recommendation data set,and the experimental results show that the model proposed in this paper has good performance in the evaluation indexes such as knowledge graph feature representation and recommendation accuracy determination.These algorithms can solve some problems existing in the current process of knowledge graph fusion into depth recommendation,and play a good role in the development of depth recommendation model,which has certain theoretical significance.