在許多現實世界的決策系統中，深度學習模型的訓練通常假設資料分佈 在部署期間將保持不變。然而，現實世界的情況是資料分佈通常會不一 樣。得到負面結果的使用者通常會做出改變使他的結果變好，這種行為 被稱為「Recourse」。隨著使用者調整行為以追求更佳結果，資料分佈也 隨之改變。當模型在此改變後的資料上重新訓練時，便會產生一個持續 的反饋迴圈：使用者行為改變模型，而更新後的模型又進一步塑造未來 的使用者行為。 本研究探討此類反饋迴圈的長期影響，分析Recourse與模型更新在時間 演進下的交互作用。透過理論與實驗分析，我們發現使用者的Recourse action 會逐漸推動模型採取更嚴格的決策標準。這一趨勢提高了使用者 實施Recourse action 的難度與代價，並使Recourse action變得較不 可靠。我們的研究結果凸顯了使用者與模型之間的互動如何演變，並在 無意間形成阻礙正面結果的高漸性門檻。;In many real-world decision-making systems, deep learning models are trained under the assumption that the data distribution will remain static throughout deployment. However, this assumption often breaks down when users actively re spond to negative predictions by changing their input features—a behavior known as recourse. As users adapt their behavior to improve future outcomes, they inad vertently alter the data distribution. When models are retrained on this altered distribution, an ongoing feedback loop is created: user behavior changes the model, and the updated models consequently reshape future user behavior. This work investigates the long-term consequences of such feedback loops by examining how repeated recourse actions and model updates interact over time. Our theoretical and empirical analysis shows that user recourse behavior gradually pushes models toward stricter decision standards. This trend increases the diffi culty and cost for users seeking recourse and makes recourse actions less reliable. Our findings highlight how user-model interactions can evolve and unintentionally create rising barriers to positive outcomes.
