Dynamic Binning for the Unknown Transient Patterns Analysis in Astronomical Time Series
Thanapol Phungtua-Eng, Yoshitaka Yamamoto, Shigeyuki Sako · 2021 IEEE International Conference on Big Data (Big Data) · 2021
In recent years, there arises a new opportunity for discovering transient phenomena such as supernovae, solar flares, and bursty events from detecting unknown transient patterns in astronomical time-series data. However, since these transient phenomena usually happen with unpredictable characteristics in shapes, sizes, and durations, scientists might lose some significant information due to the huge volume of astronomical data to be analyzed. Data sketching is useful to deal with such huge time-series data. A simple sketching technique is known as binning that captures the statistical summary of each bin of data points. In this paper, we attempt to provide a novel framework of data sketching for a statistical hypothesis testing and apply it for unknown transient pattern detection. The principal idea of statistical hypothesis testing lies in that two short-term and similar bins are mergeable into a long-term bin. By applying our proposed method, we suppress the unnecessary data while keeping the primary information without setting the bin size in advance. We evaluate our proposed method through experiments on the light curves in real-world data from telescopes with synthetic mixed-type transient patterns. Experimental results demonstrate that our proposed method outperforms several frameworks of transient pattern detection in astronomy.