PoliCE: 利用策略網路高效生成基於通用因果模型的反事實解

Abstract

近期於「反事實解釋」(Counterfactual Explanation, CE) 研究探索了在保持變數因果關係的狀況下，改變分類器輸出的輸入的變數過程。我們進一步研究如何在必須變動內部變數以改變分類器輸出時，保留變數間的因果關係。具體而言，我們提出了一種名為 PoliCE 的基於強化學習的演算法，透過迭代生成跨越決策邊界所需要的每一步 (調整變數的動作)。PoliCE 找出每個內部變數在父變數給定時的可變動性，並將對其的變動分解為主動變動和固有因果效應。此外，它保證了對分類器的少量存取，因此在保留特徵因果關係的同時，可以非常高效。實驗結果顯示，PoliCE 在包括數值和類別變數的合成和真實數據集上，比過去的方法在多項指標中表現更好，尤其在保持變數間的因果關係及效率上提升顯著。

Recent studies of Counterfactual Explanation (CE) explore the perturbation process of input features to change a classifier’s output in awareness of the causal relations among features. We further study how to preserve the inherent feature causality when the perturbation on endogenous features is necessary to changing the classifier output. Specifically, we propose PoliCE, a reinforcement learning-based algorithm to iteratively generate every step (action of tuning features) along the way to crossing the decision boundary. PoliCE finds out the perturbability of each endogenous feature given its parent features and decomposes the perturbation on it into active action and inherent causal effect. It guarantees a small number of accesses to the classifier, thus making it very efficient while preserving the feature causality. Extensive experiment results show that PoliCE outperforms the baselines on both synthetic and real datasets with both numerical and categorical features, especially in causality preservation and efficiency.