An Enhanced Detection of Outlier using Independent Component Analysis among Multiple Data Instances via Oversampling

Abstract

Anomaly is a pattern of data that does not conforms to expected behavior. It is also referred as outlier, exceptions, peculiarities, surprise etc. Anomaly detection aims to identify a small group of instances which deviates from the existing data. It needs to solve an unsupervised yet unstable data learning problem.Detecting an anomaly is an essential research topic in data mining to solve the real world applications like intrusion detection, homeland security to identify the deviated data instances. Mostly anomaly detection methods are implemented in batch mode it requires more computation and memory. Existing system online oversampling Principal Component Analysis (osPCA) algorithm to address this problem and for detecting the presence of outliers from a large amount of data via an online updating technique. In PCA normal data with multi clustering structure and data is in an extremely high dimensional space is not supported. It is typically not easy to use linear models such as PCA to estimate the data distribution if there exists multiple data instances. To overcome these problems and support multiple data instances we proposed a system called Independent Component Analysis (ICA) in which it is a technique of array processing and data analysis aiming at recovering unobserved data samples from oversampled dataset and it is also used to reduce the computation and memory requirement for outlier detection.