穩定與不確定性感知的統計機器學習於局部通用模型適用可解釋人工智慧

Abstract

隨著機器學習模型在醫療、金融和刑事司法等高風險領域的廣泛應用，對於具有穩健性、可解釋性和可信度的模型解釋需求日益增加。然而，傳統的可解釋人工智慧方法，如局部可解釋通用模型適用解釋技術，假設決策邊界在局部範圍內呈線性，並依賴基於擾動的採樣方式，導致解釋結果的不穩定性、局部擬合度低，以及特徵貢獻評估的偏差，特別是在非線性決策邊界的情況下。為了解決這些挑戰，我們提出了一系列改進的可解釋人工智慧方法，以提升局部擬合度、穩定性及不確定性量化能力，同時促進機器遺忘技術的發展。 首先，我們提出一種新的非線性局部解釋方法，該方法結合加權多元自適應迴歸樣條與自助法聚合技術，不僅能更準確地捕捉非線性決策邊界，還能在貝氏框架內建模特徵貢獻的不確定性。我們的實驗結果顯示，該方法在局部擬合度和穩定性方面表現優越，並且在多種真實世界與模擬數據集上均展現出穩定的表現。 其次，我們提出改進版的局部可解釋方法，解決了傳統技術因為擾動樣本偏離真實數據分佈而導致的不穩定性問題。該方法透過基於近鄰與核密度建模的技術，使用內部分佈樣本進行擾動，從而提升局部擬合度和穩定性，並增強對抗攻擊的韌性。我們的實驗結果表明，這種改進方法能有效減少錯誤或偏差解釋，確保更可靠的模型可解釋性。 最後，我們將上述可解釋技術應用於機器遺忘，以實現刪除訓練後模型中錯誤數據影響之能力。本研究提出一個整合式框架，結合 反事實隱性不確定性解釋與認證刪除方法，並進一步以監督式變分自編碼器搭配預測重標註機制強化反事實生成品質。此架構可先透過不確定性感知的反事實樣本辨識訓練資料中有害樣本，並透過近似貝氏推論原則實現高效且穩健的資料移除與模型修復。更重要的是，該方法不僅能移除誘發不確定性的資料，亦能透過可信任的反事實樣本進行學習，從而同時提升模型之泛化能力與可解釋性。 通過提升局部擬合度、穩定性、對抗攻擊韌性，以及將不確定性感知技術應用於機器遺忘，我們的研究推動了更加可解釋且可靠的機器學習模型發展，促進其在關鍵應用領域的實際部署。

As machine learning (ML) models increasingly permeate high-stakes domains such as healthcare, finance, and criminal justice, the need for robust, interpretable, and trustworthy explanations intensifies. Traditional eXplainable Artificial Intelligence (XAI) methods, such as Local Interpretable Model-agnostic Explanations (LIME), assume local linearity and rely on perturbation-based sampling, which leads to instability, low fidelity, and biased feature attributions—particularly in non-linear decision boundaries. To address these challenges, we propose novel XAI methods that improve local fidelity, stability, and uncertainty quantification while also facilitating machine unlearning. First, we introduce BMB-LIME (Bootstrap aggregating Multivariate adaptive regression splines Bayesian LIME), a nonlinear local explainer leveraging weighted multivariate adaptive regression splines (MARS) with bootstrap aggregating. BMB-LIME not only captures non-linear decision boundaries more effectively than existing methods but also models the uncertainty of feature contributions within a Bayesian framework. Experimental results demonstrate its superior local fidelity and stability across diverse real-world and simulated datasets. Second, we present KDLIME (KNN-kernel Density-based perturbation LIME), an improved variant of LIME that addresses instability caused by out-of-distribution perturbations. By employing KNN-kernel density-based perturbation using in-distribution samples, KDLIME achieves higher local fidelity and stability while enhancing robustness against adversarial attacks. Our empirical evaluations confirm that KDLIME mitigates the risk of producing misleading or biased explanations, ensuring more reliable model interpretability. Lastly, we extend these advancements in XAI to the domain of machine unlearning—a critical process for eliminating adversarial or erroneous data influences from trained models. We propose a unified framework that integrates Counterfactual Latent Uncertainty Explanations (CLUE) with Certified Removal (CR), and further enhances CLUE via a supervised variational autoencoder (SVAE) equipped with predictive relabeling. In this framework, CLUE identifies harmful training instances through uncertainty-aware counterfactuals, while CR performs principled and efficient unlearning through approximate Bayesian inference. By simultaneously removing uncertainty-inducing data and learning from low-uncertainty, relabeled counterfactuals, the method improves both the model's generalization and interpretability—ultimately contributing to more transparent, robust, and trustworthy ML systems. By addressing local fidelity, stability, adversarial robustness, and unlearning within an uncertainty-aware XAI framework, our research contributes to the development of interpretable and reliable machine learning models, advancing their deployment in critical applications.