一個基於動態資料驅動應用系統的動態權重多數決演算法-以概念飄移為例

Abstract

在資料不斷變動的環境中如何有效的了解資料間的相關聯性並避免預測結果因時間變動而產生的概念飄移是個困難且耗時的問題。為了能在不同時間驅動特定高度相關聯性的資料來提升演算法預測效率並偵測概念飄移，因此本研究提出一個基於動態資料驅動系統(Dynamic Data-Driven Application System)的動態權重調整多數決(Dynamic Weighted Majority)演算法。藉由所提出的演算法利用動態權重多數決來了解資料關聯性且將即時回饋至演算法，從而提供一個適用於動態資料環境下的資料預測的解決方案並同時可以偵測概念飄移的發生。本研究利用模擬資料與真實颱風(海棠颱風)歷史資料來驗證，模擬結果證實本文所提出的基於動態資料驅動系統的動態權重調整在模擬資料的預測結果上可達到89%的準確度，在真實資料的預測準確度可達到90%，並且可以發現出概念飄移的現象。

In a dynamic environment, data are changed almost instantly. It is difficult and time-consuming to find the correlations between data. At the same time, concept drift might happen along with data change in the dynamic environment. In order to stimulate the highly correlated data to support better prediction and detect concept drift, this thesis proposes a dynamic weighted majority (DWM) algorithm based on distributed dynamic data-driven Application system (DDDAS) to solve the issue. The proposed algorithm tries to find the correlations between data by DWM. Moreover, it is capable of detecting concept drift. Both simulation data and real world (HAITAUNG typhoon) data are used to validate the proposed algorithm. The result show the proposed has up to 89% accuracy in simulation case and have 90% accuracy in real world case.