Anomaly Detection for Highly Imbalanced Data–an Empirical Analysis
Akshat Ajay Das, Veena Mayya, M. M. Manohara Pai · 2023
An event or an observation that is statistically different from the others is termed an anomaly. Anomaly detection is the process of identifying such anomalies. Anomaly detection is an effective tool for risk mitigation, fraud detection, and improving the system's robustness. It is also an active research area, with numerous algorithms being proposed. In this paper, we compare the performance of various anomaly detection algorithms on mul-tivariate as well as univariate datasets. The assessment measures generated are important and can be beneficial for predicting anomalies in a timely and accurate manner. Experimental results demonstrate that on a univariate dataset, the auto-regressive moving average (ARMA), performs better than the local outlier factor (LOF), while on a multivariate dataset, the LOF model performs better. The prototype developed has been extensively tested on publicly available datasets and can be evaluated on larger, more comprehensive datasets for deployment in the real-time anomaly detection setup.