基於自動編碼器之重建值判別學習應用於異常值偵測

Abstract

異常值檢測目的在於從數據集中找出屬性值與正常資料點與不同的 離群資料點。隨著資料演化快速發展，在分析龐大數據上需要更有效 的異常值檢測，而自動解碼器是一個可以有效偵測離群值的工具，透 過重建時的誤差，以離群值具有相對正常資料點更大的誤差來判斷。 然而，重建誤差在一些數據集中會嚴重重疊，導致預測不精確。 本論文提出一個基於自動編碼器的改良解決此問題。訓練過程中， 在週期的迭代，透過閾值決定每個實例的正和負類別形成一個預測向 量，並將預測向量作為區分資訊加入到下一次迭代做成本運算。利用 區別性學習，正常資料點和離群值更可透過重建誤差做分離，使得重 建誤差成為更好的預測判別指標，進而提升異常值偵測的效能。最後 將提出的方法應用在三個常用於異常值偵測研究的數據集進行測試， 實驗結果顯示所提出的方法可以在偵測異常值達到較高的準確度。

Outlier detection aims to find the instances that are very different from the defined normal instances in a given dataset. Autoencoders are effective tools for outlier detection by utilizing the reconstruction errors, that is, the outliers have relatively larger reconstruction errors than the inliers. Nevertheless, the reconstruction errors will overlap significantly in some dataset, which leading to inaccurate prediction. In the thesis, we propose a modified autoencoder to solve the problem. Based on the autoencoder, we assign a positive and negative label to each instance and feed the prediction vector to the next iteration as a discriminative information in the learning process periodically. With the discriminative learning, the reconstruction errors of inliers and outliers are more separable, leading to a more accurate outlier detection. We have tested on three datasets that are widely used for outlier detection: Ionosphere, Wisconsin breast cancer and NSL-KDD. The proposed approach can achieve 94.30%, 97.07%, and 92.74% accuracy respectively. The experimental results show that our approach can reach high performance on identifying outliers.