Outlier Detection based on Transformations for Astronomical Time Series

In this work we propose an outlier detection method for astronomical light curves, based on time series transformations as temporal shift, masking, warping, etc. It is assumed that the outliers are unknown, and that we have access only to a set of inliers. A neural network encoder is used to learn a representation of a light curve minimizing the distance between objects of the same class and maximizing it otherwise. Each light curve is encoded as a single vector. An outlier score is computed based on the distance to the nearest class centroid. The model is applied to datasets from the Zwicky Transient Facility (ZTF), All Sky Automated Survey (ASAS) and Lincoln Near-Earth Asteroid Research (LINEAR) surveys. For model selection, surrogate metrics are estimated with the validation set. The metrics under test are the average hit ratio of the k-nearest neighbors of each light curve in the representation space, silhouette coefficient, Calinski-Harabasz index and Davies-Bouldin index. The results show that the proposed model outperforms state-of-the-art methods based on time series features and neural network approaches, reaching an average AUCPR of 0.89 for detecting outliers in the three datasets.