利用混合模型決定錨變數

Abstract

在機器學習中，我們常假設測試資料的分布與訓練資料相同。然而，當訓練分布與測試分布之間發生分布轉移（distribution shift）時，模型的預測效能可能會下降。為了解決這個問題，有一種稱為「Anchor Regression」的方法被提出，並在某些情境中展現出處理分布轉移的能力。在應用此方法時，需指定一個anchor變數，用以捕捉不同資料集之間的異質性。然而，在實務上辨識出資料集間真正潛在的差異並不總是可行的。因此，本文提出了一些用於決定anchor變數的潛在方法。這些方法主要利用非監督式學習技術以萃取具意義的特徵，並透過模擬與實際資料集進行評估，以檢驗其有效性。

In machine learning, we often assume that the testing distribution is the same as the training distribution. However, when a distribution shift occurs between the training and testing distributions, the model’s predictive performance may decline. To address this issue, a method called Anchor Regression has been proposed, demonstrating its ability to handle distribution shifts in certain scenarios. When applying this method, it is essential to specify an anchor variable, which captures the heterogeneity across different datasets. However, identifying the true underlying differences between datasets is not always feasible. Therefore, this paper proposes potential methods for determining the anchor variable. These approaches primarily leverage unsupervised learning techniques to extract meaningful features, followed by evaluations through simulations and real datasets to assess their effectiveness.