Reconstruction probability-based anomaly detection using variational auto-encoders

Anomaly detection is a method of categorizing unexpected data points or events in a dataset. Variational Auto-Encoders (VAEs) have proved to handle complex problems in a variety of disciplines. We propose a technique for detecting anomalies based on the reconstruction probability of VAEs. The proposed method trains VAEs on three different datasets. The reconstruction probability is a much more principled and realistic anomaly score than the reconstruction error utilized by auto-encoders and other data compression methods because of the theoretical background and by including the concept of variability. The paper describes recent deep learning models for anomaly detection, as well as a comparison to other methodologies. Variational auto-encoders are trained on three different datasets, in an unsupervised setup to classify the anomalies, based on reconstruction probability. Further, the in-depth study of anomaly detection techniques is presented in this paper. The data are reconstructed using the VAEs generative characteristics to investigate the root cause of the anomalies.