深度證據分位數迴歸模型結合飄移偵測器在線上學習之應用

Abstract

本研究提出一DEQPI（Deep Evidential Quantile Prediction Interval）的線上學習架構，結合深度證據分位數迴歸（DEQR）模型與自適應飄移偵測器（AdaPH-Test），用以解決資料流中概念飄移、不確定性評估困難、模型更新遲滯等關鍵問題。研究動機來自於高頻資料流環境下，傳統深度學習模型在資源需求、更新速度與適應變化能力上的不足，難以勝任即時預測與決策應用。 本論文的貢獻有三：1) 在模型架構方面，採用 DEQR 預測上下分位數以構建預測區間，跳脫常態分布假設，提升對非對稱與厚尾資料的適應性與預測解釋力。2) 在飄移偵測上，提出改良型 Page-Hinkley Test（AdaPH-Test），可根據資料變化自動調整參數，提升在不同資料尺度下的泛用性與反應速度。3) 整合多種學習率調節策略，依據不確定性與預測誤差動態調整學習率，並引入正則項L^U解決高不確定性區域（HUA）中的梯度消失問題，提升模型穩定性與收斂效率。 實驗設計涵蓋模擬資料與公開資料集，結果顯示 DEQPI 架構在預測準確度、不確定性區間穩定性與飄移偵測能力上，皆優於現有線上學習方法如ARF與DEAPI，並成功應對突發與漸進式概念飄移。綜合而言，本研究所提方法在處理高頻資料流與概念飄移的挑戰下，提供一套結構簡潔、可即時反應且具有高度可解釋性的預測模型，對智慧製造、金融預測等即時應用場景具有高度實務價值。

This thesis proposes an online learning framework named DEQPI (Deep Evidential Quantile Prediction Interval), which integrates the Deep Evidential Quantile Regression (DEQR) model with an adaptive drift detection mechanism (AdaPH-Test). The framework addresses critical challenges in streaming data environments, including concept drift, difficulty in uncertainty quantification, and delayed model adaptation. The motivation stems from the limitations of conventional deep learning models in high-frequency data streams, particularly their high computational demands, slow update capabilities, and poor responsiveness to distributional changes, which hinder their applicability in real-time prediction and decision-making scenarios. This thesis makes three primary contributions: 1. Model architecture: DEQR is utilized to directly predict upper and lower quantiles, thereby constructing predictive intervals without relying on the normality assumption. This enhances the model’s flexibility and interpretability in handling asymmetric and heavy-tailed data distributions. 2. Drift detection: A modified Page-Hinkley Test, termed AdaPH-Test, is introduced to dynamically adjust detection parameters based on streaming data behavior. This adaptation improves the detector’s generalizability and responsiveness across varying data scales. 3. Adaptive learning: Multiple learning rate adjustment strategies are integrated to dynamically tune the learning rate based on uncertainty measures and prediction errors. In addition, a regularization term L^U is incorporated to resolve the vanishing gradient problem in high uncertainty areas (HUA), thereby improving model stability and convergence efficiency. The experimental evaluation, conducted on both synthetic and public datasets, demonstrates that the proposed DEQPI framework outperforms state-of-the-art online learning methods such as ARF and DEAPI in prediction accuracy, uncertainty interval reliability, and drift detection effectiveness. Moreover, it shows robust performance in both sudden and gradual concept drift scenarios. Overall, this work delivers a lightweight, interpretable, and responsive prediction model well-suited for real-time applications in domains such as smart manufacturing and financial forecasting, offering significant practical value in dynamic and uncertainty-prone environments.