Paper

“A Machine Learning Technique to Identify Transit Shaped Signals” by Thompson et al.

http://arxiv.org/pdf/1509.00041.pdf

Summary

Supervised machine learning is used to automatically identify expolanets from Kepler light curves.

Figures

Figure 1 shows some transit-like TCEs, folded on the left, and binned on the right. Figure 2 shows the same for non-transit-like TCEs.

Figure 4 shows a histogram of log(T_LPP) for (non-)transit-like TCEs.

Figure 5 shows a histogram of log(T_LPP) for injected transits, divided into pass/fail.

Procedure

Binning

(Summarized in Section 3.2, page 3, 3rd paragraph in the right column)

• Each star’s light curve is folded on the Threshold Crossing Event (TCE) periods provided by the NExScI archive
• An equal number of bins near the TCE are chosen for each light curve
  • Bins are selected such that 51 lie within, and 90 outside the transit
    • (See Section 3.5 on page 5 for details)
• The mean of each bin is used as the magnitude
• Points are sorted by phase
• Points are normalized such that the minimum occurs at -1

Dimensionality Reduction

(Summarized in Section 3.3, beginning of page 5)

• Locality preserving projections (LPP) projects the binned light curves into a lower dimensional space
  • Algorithm is similar to PCA
    • Less sensitive to outliers
    • Better at preserving locality (hence the name) for methods like k-NN
  • produces a 20-dimensional feature vector for each event

Classification

(Summarized in Section 3.5, pages 5 and 7)

• The k-nearest neighbor (k-NN) algorithm is used to label the events as transit-like or not
  • Training data are known transit-like events from kepler
    • (Section 3.4, page 5)
  • k = 15
  • Distance is measured by the Euclidean 2-norm
  • T_LPP is the mean of the k distances
    • If T_LPP for a given TCE is within the range of values for known transit-like events, it is labeled as one

Results

• Removes over 90% of non-transiting candidates from Kepler data, and retains over 99% of known transits
• Loses 1% of injected transits
  • (Injection described in Section 3.6, page 7)