A Link Prediction Method Based on Knowledge Graph Embedding

Abstract

The invention discloses a link prediction method based on knowledge graph embedding, aiming to realize fast link prediction of large-scale knowledge graph. The technical solution is to first partition the knowledge graph; then build an embedded model and a knowledge graph link prediction system composed of N servers and a shared file system. During training, the server serves as the master node and training node, and the server serves as the query node and prediction node during training. The master node is installed with a lock server process, and the training node is installed with a data loading process and a GPU training process; then multiple machines are parallelized and the CPU and GPU parallelize the embedded model for distributed training; finally N servers load the trained embedded model, and the knowledge Graph links are used for parallel prediction; the invention not only speeds up the training and connection prediction of knowledge graph embedding, but also solves the problem of embedding performance degradation caused by partitioning, and can quickly obtain high-quality knowledge graph embedding.

Description

technical field [0001] The invention relates to the field of knowledge map link prediction, in particular to a method for link prediction based on knowledge map embedding. Background technique [0002] A knowledge graph (Knowledge Graph, KG) is a structured representation of real-world information, usually representing a multi-relational graph, that is, a graph containing multiple types of nodes and relationships. KG consists of three parts. The entity set E is a collection of things in the real world, such as people, place names, concepts, drugs, companies, etc., for example, "Beijing" is an entity; the relationship set R is a certain relationship between entities. Connections, such as the "capital" relationship means that one entity is the capital of another entity; the triple set T represents the fact that there is a certain relationship between entities, such as the triple (Beijing, capital, China), which means "Beijing is the China's capital" fact. [0003] Although k...

AI Technical Summary

Problems solved by technology

[0012] However, the PBG knowledge graph embedding framework has the following problems: 1. During partition training, the server first uses the CPU for data exchange and preprocessing, and then loads the embedding and triples to the GPU for calculation. These two processes are serial , so the CPU and GPU need to wait for each other to increase the entire training time; 2. Because of the partition, the data distribution of the partitioned triple block and the data distribution of the entire knowledge graph triple are very different, so the final embedding The performance of the model will decrease. The experimental results show that when performing link prediction, MRR (Mean ReciprocalRank) decreases as the number of partitions increases; 3. Only embedding training and testing of knowledge graphs are provided, lacking knowledge Link Prediction Applications with Graph Embedded in Distributed Clusters

Images