Knowledge Graph Reasoning Model Based on Negative Sampling Optimization

In the process of training a network security knowledge graph reasoning model, a large number of negative samples are required to enhance the model's generalization ability and accuracy. Considering that traditional negative sampling methods fail to generate high-quality negative samples, which in turn affects the accuracy of the network security knowledge graph reasoning model, based on Knowledge graph embedding by translating on hyperplanes (TransH), we propose an improved negative sampling TransH model (INS-TransH). This model constrains the negative samples based on the syntactic rules between entities and relations, thereby further enhancing the quality of generated negative samples. It also applies a self-adversarial negative sampling method that calculates the scores of negative samples to filter out low-quality samples. We construct a network security knowledge graph using Common Vulnerabilities and Exposures (CWE) database, Common Weakness Enumeration (CVE) database, and Common Attack Pattern Enumeration and Classification (CAPEC) database, and use this knowledge graph as the dataset for the INS-TransH model. Experimental results demonstrate that the INS-TransH model exhibits improved knowledge reasoning capability, with enhancements in Hits at 10 (Hits@10) and Mean Reciprocal Rank (MRR) metrics.