Exoplanet Validation

Problem

When searching for new planets through transit detection in NASA's Kepler satelite data, a significant portion of time is spent on validation of the detected signal via:

• Slow Monte-Carlo validation methods
• Blender (Torres et al. 2010)
• Vespa (Morton & Johnson 2011)
• Pastis (Diaz et al. 2014)
• Time consuimg follow-up observations

In this work:

• Optimise and test ML methods in classying planetary signal into False Positive or Confrimed
• Evaluate the quality of predicted False Positive probabilities and compare to existing validation methods
• Vespa (Morton & Johnson 2011)

Pre-trained models would allow to:

• Quickly validate planets, and only focus on uncertain signals
• Save on follow up observation resources

Preprocessing & Data

We follow procedures in McCauliff et al. (2014) and use:

• Thershold Crossing Events (TCE) catalog
• 34,024 detected TCEs
• KOI catalog (Burke et al. 2014)
• Known, already validated planets

Matching two catalogs gives us 4049 instances for training:

• 2238 Confirmed planets
• 1810 False Positives
• 1189 unlabeled signals
• 139 features per instance

Two additional statistics computed:

• Maximum ephemeris correlation (McCauliff et al. 2014)
• Self Organising Map statistic (Armstrong et al. 2016)

Feature Importance

Feature Importance (following McCauliff et al. 2014):

1. Fit Random Forest model
2. Permute an attribute within out-of-bag data of the tree
3. Predict and compute the error

Mean increase in error is the importance of the permuted feature

Correlation

1. Compute Pearson's correlation between each attribute
2. If correlation above threshold, remove least important attribute

Methods and Results

• Simple decision boundaries are sufficient
• Decision Tree ensembles are "to good" for us
• Investigate the normality of each point using properties of Random Forests (Liu et al. 2008)

For each instance $i$, compute normality score $N_i$

$$N_i = 2^{\frac{\mathbf{E}(D_i)}{\mathbf{E}(P_i)}}$$