概念飄移下基於少數類別導向局部上採樣之不平衡資料流線上學習架構

Abstract

在半導體製造、設備監控、異常偵測與金融詐欺偵測等即時決策場域中，資料通常以連續到達的資料流形式產生，模型需即時預測並持續更新。然而，實際資料流常同時面臨概念飄移與類別不平衡兩項挑戰。當兩者並存時，模型更新過程容易受多數類別樣本主導，使少數類別決策區域難以及時修正，進而增加少數類別漏判風險。此時，整體準確率仍可能維持於較高水準，因而掩蓋少數類別辨識能力的退化，使關鍵樣本漏判情形不易被察覺。 為解決上述問題，本研究提出少數類別導向局部上採樣之線上學習架構（Minority-Guided Local Oversampling Online Learning, MILO），以處理概念飄移與類別不平衡並存下的資料流分類問題。所提出架構以證據式集成學習（Evidential Ensemble Learning）作為分類與不確定性量化基礎，透過多個證據式深度學習（Evidential Deep Learning）模型取得預測資訊，並由元模型整合各證據式深度學習模型之輸出，以產生最終分類結果；同時，該架構亦可取得資料不確定性與模型不確定性資訊，以支援非平穩資料流中的預測可靠性評估與模型狀態觀察。 在線上更新過程中，本研究設計少數類別訓練權重動態調整機制，使模型能依據近期資料狀態與模型表現，動態提高少數類別樣本於訓練過程中的影響程度。同時，本研究以少數類別漏判情形作為局部樣本補強依據，於漏判樣本之鄰近特徵空間進行少數類別導向局部上採樣，並利用所生成之樣本進行模型再訓練，以強化少數類別資訊對模型更新的影響，使模型能隨資料流變化持續修正決策邊界。 本研究以合成與公開資料集評估所提方法，並與代表性線上學習方法及不同模型設定進行比較。實驗結果顯示，MILO透過訓練權重動態調整與局部樣本補強，可提升少數類別辨識能力，並提供分類預測與不確定性量化資訊，顯示其於不平衡資料流環境中的適用性。

In real-time decision-making applications such as semiconductor manufacturing, equipment monitoring, anomaly detection, and financial fraud detection, data typically arrive continuously as streams, requiring models to make predictions and be continuously updated in real time. However, real-world data streams often face the dual challenges of concept drift and class imbalance. When these challenges coexist, model updates can easily become dominated by majority-class samples, making it difficult to promptly adjust the minority-class decision region and consequently increasing the risk of minority-class false negatives. Meanwhile, overall accuracy may remain relatively high, thereby masking deterioration in minority-class recognition and making missed critical instances difficult to detect. To address these issues, this study proposes Minority-Guided Local Oversampling Online Learning (MILO), an online learning framework for data stream classification under concurrent concept drift and class imbalance. The proposed framework adopts Evidential Ensemble Learning (EEL) as the basis for classification and uncertainty quantification. Multiple Evidential Deep Learning (EDL) models are used to obtain predictive information, while a meta-model integrates the outputs of the individual EDL models to produce the final classification result. In addition, the framework provides estimates of data and model uncertainty to support prediction reliability assessment and model-state monitoring in non-stationary data streams. During online updates, this study introduces a dynamic adjustment mechanism for minority-class training weights, enabling the model to increase the influence of minority-class samples during training according to recent data conditions and model performance. Meanwhile, minority-class false negatives are used as the basis for local sample augmentation. Minority-guided local oversampling is performed in the local feature space surrounding the false-negative samples, and the generated samples are subsequently used to retrain the models. This process strengthens the influence of minority-class information on model updates and enables the decision boundary to be continually adjusted as the data stream evolves. The proposed method is evaluated using synthetic and public datasets and compared with representative online learning methods and different model configurations. The experimental results show that MILO can improve minority-class recognition through dynamic training-weight adjustment and local sample augmentation, while also providing classification predictions and uncertainty estimates, demonstrating its applicability to imbalanced data streams.