Conformal Kernel Expected Similarity for Anomaly Detection in Time-Series data

Abstract

The problem of anomaly detection arises in many practical applications. Currently it
is highly important to be able to detect outliers in data streams, as recent years have seen a rapid
growth in the amount of such data. Only a few techniques are applicable to real-time data and even
fewer could provide an interpretable anomaly score. Probabilistic interpretation of the anomaly
score could allow an analyst to choose the anomaly threshold based on the desired false alarm rate,
which is highly important in a number of real-life applications. We propose a modification of the
EXPoSE algorithm for anomaly detection in time series data, which produces a probabilistic score
of abnormality. The proposed algorithm is developed within the framework of conformal anomaly
detection and utilizes the expected similarity as a measure of non-conformity.