Features

For the task of classifying these lightcurves we will be using a series of 42 features which can be derived from the raw time series data, without requiring specific lightcurve lengths or time resolutions. When extracting features for classifying against different phenomena we will necessarily require a variety of different time bin lengths, due to the temporal natures of the different phenomena.

These features are also selected in order to attempt to reduce the computational load during their calculation, such that they can be determined quickly for detections in new observations on an ongoing basis. Where definitions of features are in agreement with those defined in specific papers, or have methodologies described in specific papers those will be appropriately referenced. The features are listed thus:

• Proportion of time bins where the count rate is greater than mean - 5/4/3/2/1 * sigma (Features 1-5)
• Proportion of time bins where the count rate is greater than mean + 1/2/3/4/5 sigma (feat. 6-10)
• Proportion of time bins where the count rate is within mean +/- 1/2/3/4/5 sigma (feat. 11-15)
• Interquartile range / sigma (feat. 16)
• sigma / range (feat. 17)
• IQR / range (feat. 18)
• Proportional position of the LQ/Median/UQ within the range (feat. 19-21)
• LQ/Median/UQ - mean / sigma (feat. 22-24)
• Maximum absolute difference between consecutive points / sigma (feat. 25)
• Kurtosis (feat. 26)
• Skew (feat. 27)
• Robust median statistic (feat. 28) ^
• Median Absolute Deviation / sigma (feat. 29)
• Reverse Cross-correlation normalised by sigma (feat. 30)
• Autocorrelation peak beyond the first zero (feat. 31)
• Consecutive same-sign deviation proportion (feat. 32) ^
• Consecutive same-sign difference deviation proportion (feat. 33)
• Lomb-Scargle periodogram peak (feat. 34) **
• Reduced chi-square against a constant value (feat. 35)
• Normalised excess variance (feat. 36) ^
• Number of peaks from Bayesian block analysis / number of data points (feat. 37) @
• Amplitude of largest Bayesian flare / sigma (feat. 38)
• Anderson-Darling test (feat. 39)
• S_B variability detection statistic (feat. 40) ^
• Von Neumann ratio (feat. 41) ^
• Excess Abbe value (feat. 42) ^

---

References

^ Comparative performance of selected variability detection techniques in photometric time series, Sokolovsky, K.V., et al., 2017, https://ui.adsabs.harvard.edu/abs/2017MNRAS.464..274S/abstract

** Least-Squares Frequency Analysis of Unequally Spaced Data, Lomb, N.R., 1976, https://ui.adsabs.harvard.edu/abs/1976Ap%26SS..39..447L/abstract

@ Studies in Astronomical Time Series Analysis. V. Bayesian Blocks, a New Method to Analyze Structure in Photon Counting Data, Scargle, J., 1998, https://ui.adsabs.harvard.edu/abs/1998ApJ...504..405S/abstract