High Dimensional Time Series Anomaly Detection

Abstract

Anomaly Detection task is to determine critical data points whose behaviour deviates unexpectedly from usual data points behaviour. These anomalous data points might indicate a major fault in a manufacturing unit, security glitch on a server, fraud in banking system or abnormal functioning of a human body part. These data are usually recorded using several high precision sensors to capture multiple contributing factors as multivariate time series data. To determine the anomalous data on such high dimensional real world data using data driven machine learning approaches, it is important to extract temporal information among data points and latent feature based methods are instinctive choice. In this thesis primarily, we will propose an Anomaly Aware Matrix Factorization (ATMF) method with two temporal neighborhood term , the autoregressive bias which could learn local patterns and the moving average bias, which could smoothen noises. To develop an optimization function which will be robust to outliers we will use an approx mean absolute error function. ATMF performances will be demonstrated using 5 real world dataset. This thesis further propose future modifications in this model to encode contextual information within model using a minimum arborescence tree. In the second part, we will briefly discuss, an automated time series outlier detection System (TODS) package for high dimensional time series data, which has several modules for data preprocessing, feature extractions and anomaly detection