基于主动学习和二次有理核的模型无关局部解释方法  

Local Interpretable Model-agnostic Explanations Based on Active Learning and Rational Quadratic Kernel

作　　者：周晟昊 袁伟伟[1] 关东海 ZHOU Shenghao;YUAN Weiwei;GUAN Donghai(College of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211100,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211100

出　　处：《计算机科学》2024年第2期245-251,共7页Computer Science

基　　金：国防基础科研计划(JCKY2020204C009)。

摘　　要：深度学习模型的广泛使用,在更大程度上使人们意识到模型的决策是亟需解决的问题,复杂难以解释的黑盒模型阻碍了算法在实际场景中部署。LIME作为最流行的局部解释方法,生成的扰动数据却具有不稳定性,导致最终的解释产生偏差。针对上述问题,提出了一种基于主动学习和二次有理核的模型无关局部解释方法ActiveLIME,使得局部解释模型更加忠于原始分类器。ActiveLIME生成扰动数据后,通过主动学习的查询策略对扰动数据进行采样,筛选不确定性高的扰动集训练,使用迭代过程中准确度最高的局部模型对感兴趣实例生成解释。并且,针对容易陷入局部过拟合的高维稀疏样本,在模型损失函数中引入了二次有理核来减少过拟合。实验结果表明,所提出的ActiveLIME方法引比传统局部解释方法具有更高的局部保真度和解释质量。With the widespread use of deep learning models,people are more aware that the decision-making of model is a problem that needs to be solved urgently.Complex and difficult-to-interpret black-box models hinder the deployment of algorithms in actual scenarios.LIME is the most popular method of local interpretation,but the resulting perturbed data is unstable,leading to bias in the final explanation.To solve the above problems,local interpretable model-agnostic explanations based on active learning and rational quadratic kernel,ActiveLIME,is proposed,which makes the local interpretable model more faithful to the original classifier.After ActiveLIME generates the perturbed data,it samples the perturbation through the query strategy of active lear-ning,selects the perturbation with high uncertainty for training,and uses the local model with the highest accuracy in the iteration to generate explanations for the instances of interest.And for high-dimensional sparse samples that are prone to local overfitting,a rational quadratic kernel is introduced into model’s loss function to reduce overfitting.Experiments indicate that the proposed ActiveLIME has better local fidelity and quality of explanations than traditional local explanation algorithms.

关 键 词：局部解释 扰动采样 主动学习查询策略 二次有理核 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]