Research On Counterfactual Explanation Generation Method With Minimal Feature Boundary

The complex and opaque decision-making process of machine learning models limits the interpretability of predictions,resulting in the inability to effectively mine results beyond the empirical.Counterfactual explanation,which can track the causal mechanisms behind data,is a hot topic in the field of interpretable machine learning.However,existing methods lead to lower prediction accuracy by ignoring the interference of irrelevant and redundant features in the instances,or lead to false causal explanations due to the inability to accurately identify the key causes affecting the target variables,resulting in unstable algorithms and large reversal costs.By injecting counterfactual explanations into the prediction model and automatically adjusting feature values to generate counterfactual instances within a minimal boundary to achieve prediction inversion,the causal relationships between target variable and features can be explained more accurately.The main elements of this study are as follows.First,considering that the presence of irrelevant or redundant features in data instances can reduce the prediction accuracy and lead to false causal explanations,the min FB algorithm for mining the minimal feature boundary（MFB）is proposed.The min FB algorithm mines the Markov boundary of the target variable using conditional independence tests,and then introduces additive factors to supplement the causal features that may be missed to obtain the MFB with irrelevant or redundant features removed.Second,the counterfactual explanation generation algorithm CEG_MFBbased on the minimal feature boundary（MFB）is proposed.The algorithm uses the MFB mined by min FB algorithm as the generation range of counterfactual instances,and automatically adjusts the feature values within the MFB to achieve prediction inversion with minimum cost.By limiting the counterfactual changes to the range of MFB,i.e.,the causal features of the target variable,invalid adjustments can be reduced,reducing counterfactual generation costs and false causal explanations.Third,the performance of the algorithm is verified by parametric analysis and experimental comparison.The performance of the min FB algorithm and the effectiveness of MFB for counterfactual explanation were verified using 16 representative datasets from different domains.In addition,the performance of CEG_MFBis analyzed by comparing it experimentally with state-of-the-art counterfactual explanation generation algorithms in terms of evaluation metrics such as validity,proximity,sparsity,and distance.Finally,CEG_MFBand its comparison algorithm are applied to a real glioma classification scenario to find classification decision boundary using the obtained minimal feature boundary to explain the grading of gliomas using the generated counterfactual instances.The effectiveness of the algorithm is verified by analyzing the experimental results.