兩階段時間序列資料探勘架構：以即時股市成交量分析為例

Abstract

時間序列資料是一連串隨時間變動的資料序列。過去應用資料探勘於時間序列的分析大多強調其圖形特徵的比對及相似度的測量。其雖然可用視覺化的方式呈現所找出的特徵樣式，卻無法以簡單的以文字描述所得到的結果。本研究提出兩階段時間序列資料探勘之架構，試圖結合不同的探勘技術來處理以上的問題。此架構包含了資料探索以及關聯規則技術。研究中應用此架構對於即時股市成交量進行分析，得到可量化的關聯規則，並經由有效性分析，證實規則的通用性以及架構的可行性。最後，本研究將傳統資料探勘與本架構之分析結果進行比較，證明兩階段架構對於時間序列複雜的特性具有良好的分析能力，並能優於傳統的探勘方式。

Time series data vary with time. In the past, most of the researches had focused on the matching of feature points or measuring of the similarities. They can successfully represent the feature patterns in a visualized way. In the mean while, those researches failed in describing the results in simple and understandable words. In this thesis, a two-phase architecture for mining time series data is introduced. By combining some different mining techniques, the difficulties mentioned above may be overcome. This architecture mainly consists of Exploratory Data Analysis (EDA) and techniques related to mining association rules. After analyzing the real-time stock market trading volume, quantitative association rules are available. Meanwhile, the generality of rules and the feasibility of this architecture can be proved through accuracy analysis. Finally, an analytic result comparison between the traditional data mining techniques and this architecture is done. The comparison results show that the two-phase architecture is superior to traditional techniques in its ability of analyzing the time series data.