Research On Knowledge Graph Completion Model Combining Temporal Convolutional Network And Monte Carlo Tree Search

Network structure data such as social network,knowledge map and so on have become very important in the era of big data.However,unlike Euclidean data such as pictures and videos,staggered and complex non-Euclidean structure data surface always has the problem of feature extraction.Complex data interaction form and huge data volume make the mining and analysis of network structure data change It's relatively difficult.The traditional network structure data mining usually uses the representation learning method,that is,the nodes and relations in the network get the embedded vector,and the direction and module length of the vector have the representation significance.In addition,with the rise of reinforcement learning method,the task of applying reinforcement learning model to data mining of network structure data processing appears.In this paper,knowledge map is used to analyze and mine data from the perspective of reinforcement learning.One of the main problems is the entity completion of knowledge map.Therefore,this paper starts with the task of knowledge map completion,which is expressed as the relationship between the given initial entity and the entity,and then completes the target entity corresponding to the relationship,forming a complete(head entity,relationship,tail entity)triple.This task can be applied to many downstream tasks such as knowledge map completion,question answering system,recommendation,etc.In this paper,the completion task process is formally defined as Markov process,and reinforcement learning is used to solve the completion problem.The network feature extraction and strategy learning method can adapt to a variety of tasks.First of all,this paper defines the Markov process of knowledge map completion task in detail,and defines the four elements of the process,namely,state space,action space,transfer function and return function,so that the process can learn task objectives in the way of reinforcement learning.Then,this paper constructs a deep agent network(Graph-Agent,GA)is used to learn the action decision-making strategies in different states in the learning environment,in which a state feature extraction layer network based on convolution network in time domain is proposed for the serialization of states;a complete action space mapping and a flexible strategy layer network based on shared parameters are proposed for the uncertainty of action space in different states to form a complete depth agent in reinforcement learning Network.In the process of deep agent network environment exploration,agent decisionmaking combined with Monte Carlo tree search is used to obtain training data,which solves the problem of low return rate of upstream walk sampling of graph structure data,and uses path storage pool and separation strategy to train agents to update the completion model.Finally,in the prediction stage of the model,the Q-value similar to the exploration process and the Monte Carlo tree search method are used to rank the search results by weighted score.The experiment is carried out on the ten relational datasets and wn18 rr datasets of nell995,and uses the information retrieval evaluation algorithms such as the average reciprocal ranking(MRR),hits @ K,and the average precision mean(map)to verify the effect of the model.At the same time,the experiment is compared with the classical algorithms in the three kinds of algorithms commonly used in this task,and the interpretability of the model is analyzed to a certain extent.The training speed of the state coding network in the model is faster than that of the cyclic neural network.The experimental results show that the average map of NELL995 dataset is 89.9% in ten relational datasets,and the MRR of WN18 RR dataset is 42.5%.The evaluation of the model is close to other methods in some tasks,and the best results are obtained in multiple completion tasks.