Autocalibration of Outlier Threshold with Autoencoder Mean Probability Score
Anomaly detection is a widely studied field in computer science with applications ranging from intrusion and fraud detection, medical diagnosis and quality assurance in manufacturing. The underlying premise is that an anomaly is an observation that doesn't conform to what is considered to be normal. A problem however is in defining the threshold that draws the line between what is normal and what is an anomaly which is largely dependent on a domain expert or from empirical testing that would yield the best result. Another problem is that the availability of data with regards to what is not normal is highly unavailable in real world scenarios making it difficult for traditional machine learning techniques to build a classification model. In this study, we propose a method that automatically determines the outlier threshold using a semi-supervised learning approach with autoencoders. To validate the performance of our proposed approach, we perform several experiments in comparison with traditional outlier detection approaches as well as an existing semi-supervised approach in one class classification, specifically OneClassSVM. The goal of this study is to eventually apply the method for autocalibration of anomaly detection of frames in video sequences. Initial results are also presented in a computer vision task.
Raphael B. Alampay and Patricia Abu. 2019. Autocalibration of Outlier Threshold with Autoencoder Mean Probability Score. In Proceedings of the 2019 2nd Artificial Intelligence and Cloud Computing Conference (AICCC 2019). Association for Computing Machinery, New York, NY, USA, 104–110. DOI:https://doi.org/10.1145/3375959.3375978