使用具資料特徵位置感知的資料分布動態預測器進行概念漂移預測

Abstract

在許多實際應用中，資料以資料流的形式隨時間收集並用於訓練機器學習模型。然而，在動態變化的真實環境中，資料分布常常發生變化，這種現象被稱為概念漂移。為了應對概念漂移，先前的研究主要集中在適應模型到最新的概念上。此外，在許多情況下，存在預測驅動概念漂移的潛在因素並預測未來概念的可能性。先前的研究已經處理了針對迴歸任務的真實世界概念漂移預測和針對分類任務的漸進式概念漂移預測。然而，針對分類任務的真實世界概念漂移預測的研究仍然有限。本文提出了一種新穎的方法，名為"具資料特徵位置感知的資料分布動態預測器（LA-DP）"，旨在解決分類任務的真實世界概念漂移預測。LA-DP利用資料實例的特徵位置資訊與歷史資料分布趨勢來預測未來的資料分布。它通過採用編碼器-解碼器架構以及使用注意力機制來連接兩個模塊，展現出資料特徵位置感知的能力。我們還提出了一個框架，利用LA-DP預測的未來資料分布來生成未來的分類器。我們通過廣泛的實驗驗證了LA-DP相較於最先進方法的有效性。

In numerous real-world applications, streaming data is collected over time, and it is a common occurrence for the data distribution to undergo changes in nonstationary real-world environments. This phenomenon, referred to as concept drift, presents substantial challenges. Previous research has primarily concentrated on adapting models to the most recent concept to address concept drift. Additionally, there is often an opportunity to forecast the underlying factors driving concept drift and predict future concepts in many cases. Previous research has tackled concept drift prediction for regression tasks with real-world drift and classification tasks with incremental actual drift. However, research targeting real-world drift prediction for classification tasks is limited. In this paper, we propose the Location-aware Dynamic Predictor (LA-DP), a novel approach designed for real-world concept drift prediction for classification tasks. LA-DP leverages the location information and historical data distributions of data instances to predict future data distribution. It exhibits location-aware capability by incorporating an encoder-decoder architecture and employing the attention mechanism to establish connections between the two modules. We also develop a framework that utilizes LA-DP predictions to generate future classifiers. The effectiveness of LA-DP is demonstrated through extensive empirical experiments.