時間序列資料處理與相似性擷取

Abstract

時間序列資料的相似性擷取在很多應用中都是不可或缺的，例如：金融資料分析與金融市場預測、動態物體的搜尋與追蹤和醫學上的治療。因此，如何有效且有效率地在時間序列資料庫中，搜尋到和查詢相似的序列成為一個非常重要的議題。經由觀察，我們知道序列的形狀或趨勢在相似性上是一個決定性的特徵，為了補足現有方法當中的不足，我們需要一個用來處理序列的形狀或趨勢的新方法。這篇論文中我們提出一個等量分段線性表示法，用角度來代表時間序列資料的形狀或趨勢。以這個表示法為基礎，我們提出一個二階層的相似性測量方。第一個階層是較粗略主要趨勢的比對，也就是兩個序列的主要趨勢必須是相配的，在第二個階層是較精細比對，我們使用一個距離函數來比較兩個在第一階層配對成功的序列的距離。實驗結果顯示我們的表示法能夠正確地代表時間序列資料，而且比之前提出分段聚集近似表示法和分段線性表示法在準確率上還要優秀。實驗也顯示我們的相似性測量方式不僅是有效的，而且也是有效率的。

The similarity retrieval of time series is required for many applications, such as financial data analysis and market prediction, moving object search and tracking and medical treatment. As a result, how to effectively and efficiently search subsequence which is similar to query in time series database is an important issue. By some observations, we know the shape or trend of time series is a dominant feature of similarity. A similarity measure for dealing with shape or trend is needed. In this thesis, we propose a new representation, called Equal Piecewise Linear Representation (EPLR), to represent time series by angles. Based on EPLR, we further present a two-level similarity measure with to define our subsequence similarity. The first level is major trends match that the major trends of two subsequences must be matched. In the second level, we use a distance function to compute the distance between two subsequences which are matched in the first level. The experimental results show that our representation can correctly represents time series data and is superior to previously proposed representations, Piece Aggregate Approximation (PAA) and Piece Linear Representation (PLR), in accuracy. Besides, our similarity measure is not only effective but also efficient.