首页|基于进化集成的图神经网络解释方法

基于进化集成的图神经网络解释方法

扫码查看
针对图神经网络模型普遍缺乏可解释性问题,提出一种基于进化集成的图神经网络解释方法,为模型预测提供质量更高的解释.将当前主流图神经网络解释方法GNNExplainer和PGExplainer作为初级解释器,分别为模型预测提供初级解释;基于初级解释结果设计遗传算子,采用改进遗传算法集成两种初级解释结果得到最终解释.在4 个真实数据集和4 个合成数据集上进行广泛试验,从定性和定量两个角度对试验结果进行评估.试验结果表明,相较于同类算法,提出算法的准确度平均提高 17%,忠实度平均提高 20%.与传统集成学习融合策略相比,改进遗传算法作为集成器对解释方法的优化效果更为显著,所有指标整体平均提高 29%.采用进化集成策略能够显著提高图神经网络解释算法的性能.
Explainer for GNN based on evolutionary ensemble learning algorithm
To address the challenge of limited explainability in Graph Neural Network(GNN)models,a novel explainer employing an evolutionary ensemble learning algorithm was developed.More refined explanations for model predictions compared to existing methods were offered.Both GNNExplainer and PGExplainer,recognized as leading tools in the field,were utilized as initial explainers.Guided by the primary explanation results for the design of genetic operators,an improved genetic algorithm was employed to integrate the initial explanations into a comprehensive final explanation.Extensive experiments were conducted on four real-world and four synthetic datasets,with results evaluated from both qualitative and quantitative perspectives.The experimental results showed that the proposed approach achieved an average improvement of 17%in accuracy and 20%in fidelity over baseline methods.Compared to conventional ensemble fusion strategies,an overall average improvement of 29%across all metrics was achieved by using improved genetic algorithm as an integrator.The substantial enhancement in GNN explanation capabilities was realized through the adoption of an evolutionary ensemble strategy.

graph neural networksevolutionary algorithmsensemble learningdeep learningmachine learningexplainable AI

常新功、苏敏惠、周志刚

展开 >

山西财经大学信息学院,山西 太原 030006

图神经网络 进化算法 集成学习 深度学习 机器学习 可解释人工智能

国家自然基金青年资助项目山西省基础研究计划自然科学研究面上资助项目山西省研究生教育创新资助项目

619022262022030212212182022Y534

2024

山东大学学报(工学版)
山东大学

山东大学学报(工学版)

CSTPCD北大核心
影响因子:0.634
ISSN:1672-3961
年,卷(期):2024.54(4)