Anomaly detection has been researched as a technique to classify anomaly data that has different characteristics compared to normal observations. Distribution-based anomaly detection approaches have played a positive role among these anomaly detection...
Anomaly detection has been researched as a technique to classify anomaly data that has different characteristics compared to normal observations. Distribution-based anomaly detection approaches have played a positive role among these anomaly detection models. In general, since abnormal observations are often noticeably less frequent than normal observations, one-class classification (OCC) is employed to tackle this issue by training the model only a single class of observations representing the normal class. This thesis introduces a model for anomaly detection using autoencoders to effectively reduce high-dimensional data into a lower-dimensional space and then training only normal data. In cases where the distribution of normal observations includes skewness, setting up a symmetric distribution might mistakenly classify normal observations as anomalies. Thus, in this thesis, we aim to utilize a flexible skew-normal distribution, which is suitable for the data with strong skewness and heavy tails. Moreover, to handle data with complex distributions, we adapt the skew-normal mixture model to detect anomalies more accurately.