A method for predicting software security entity relationships based on reinforcement learning
Existing methods for entity relation prediction in translation-based software security knowledge graph lack interpretability,while those based on path reasoning exhibit low accuracy.To alleviate this issue,a reinforcement learning-based prediction method is proposed.This method first represents the structural in-formation and descriptive information of the software security knowledge graph as low-dimensional vectors us-ing the TuckER model and SBERT model respectively.Then,it models the entity relation prediction process as a reinforcement learning process,integrating the scores computed by the TuckER model into the reward function of reinforcement learning.The method further employs input entity relation vectors to train the policy network of reinforcement learning.Finally,it utilizes beam search to obtain ranked lists of answer entities and corresponding inference paths.Experimental results demonstrate that this method provides relation paths for all predicted results.In link prediction experiments(h,r,?),the hit@5 is 0.426,hit@10 is 0.797,and MRR is 0.672.In fact prediction experiments,the accuracy is 0.802,and precision is 0.916.In terms of ac-curacy,compared with similar entity relation prediction models,this method shows varying degrees of im-provement.Furthermore,through interpretability analysis experiments,the interpretability of this method is validated.
国家科技期刊平台
NSTL
万方数据
