Exoplanet Habitability Anomaly Eval

exoplanet-habitability-anomaly-eval

Postulating Exoplanetary Habitability via a Novel Anomaly Detection Method — Sarkar et al. (2021) (arXiv:2109.02273, 2021)

What this evaluates

Evaluates an unsupervised anomaly detection model's ability to identify potentially habitable exoplanets using planetary parameter features, deliberately excluding direct habitability labels to simulate real-world discovery scenarios.

Datasets

• PHL-EC (Planetary Habitability Laboratory Exoplanet Catalog) — total ?; splits: D1 (2018) (1681), D2 (Latest) (4048)

Metrics

• PHL-HEC intersection (primary) — range: count
  • Counts the number of predicted anomaly exoplanets that overlap with the ground-truth optimistic and conservative habitable exoplanet lists from the PHL-HEC catalog.

Input / output format

Input: 10 principal components derived from cleaned planetary features (continuous and categorical), with highly correlated and irrelevant features removed.

Output: Binary anomaly label (1 = anomaly/habitable candidate, 0 = normal/non-habitable) or a ranked list of exoplanets based on anomaly scores.

Scoring recipe

def score(predictions, gold_catalog):
    # predictions: list of exoplanet IDs marked as anomalies (1)
    # gold_catalog: set of exoplanet IDs in PHL-HEC optimistic/conservative lists
    overlap = len(set(predictions) & gold_catalog)
    return overlap

Common pitfalls

• Direct habitability markers like Star's Habitable Zone (HZ) and Earth Similarity Index (ESI) are explicitly removed during training to enforce unsupervised learning.
• Key features like Gravity and Density have high missingness, restricting D2 analysis to only 706 out of 4048 planets.
• Extreme class imbalance exists in D1 (1631 non-habitable vs. 2 hypopsychroplanets), which can skew anomaly detection thresholds.

Evidence (verbatim from paper)

Dataset D2: The latest version contains 4048 exoplanets and classifies planets into three classes, non-habitable planets constituting a majority (3993). The current version does not include Surface Temperature and even such important characteristics as Gravity, Density, or Escape Velocity, are available for only 706 out of 4048 of these exoplanets. Hence, only one set of inferences covering these 706 exoplanets is drawn, out of which only 6 intersect with the optimistic/conservative list of exoplanets in the PHL-HEC.

Citation

@misc{sarkar2021postulating,
  title={Postulating Exoplanetary Habitability via a Novel Anomaly Detection Method},
  author={Sarkar et al. (2021)},
  year={2021},
  note={arXiv:2109.02273}
}

• arXiv: 2109.02273