使用正規隨機漫步及相似度進行異常偵測

Abstract

資料雲幾何樹是一個透過正規隨機漫步捕捉資料結構，再進行分群的一個演算法。本論文從資料雲幾何樹的概念中延伸出了兩種異常偵測的方法，第一種是使用樣本間的相似度加總來進行異常偵測，第二種則是透過正規隨機漫步探索數據，以探索到的時間點做為異常值。而在使用多尺度的模擬資料時，發現演算法表現不穩定，因此使用了self-tuning的策略來改良演算法，能克服在資料多尺度時進行異常偵測的問題，最後在實際資料上和經典方法LOF比較。

Data cloud geometry tree is a clustering algorithm that explores data structures by regulated random walk. Based on the concept of data cloud geometry tree, the current study proposes two anomaly detection methods. The first method uses sum of similarities between samples for anomaly detection. The second method explores data through regulated random walk to detect unusual pattern. Samples that were later explored are treated as abnormal. However, the performance of the proposed algorithms are unstable when dealing with multi-scaled simulated data. Therefore, self-tuning strategy is applied to improve the performance of algorithms and to overcome the anomaly detection problem for multi-scaled data. Finally, the performance of proposed methods are compared to the performance resulting from the classical method, LOF, with many real examples.