Featured papers from Anre's Research Archive.

This study will analyze whether observable properties of stars can effectively predict the presence of gas giant planets in the same stellar systems, using supervised machine learning methods on data taken directly from the NASA Exoplanet Archive. A selective dataset of 4,119 unique, confirmed exoplanet host stars was constructed using the PSCP (Planetary Systems Composite Parameters) table, of which 1,407 (34.1%) host at least one gas giant planet, which in this experiment is defined as a planet with a measured radius greater than 0.8 Jupiter radii (an alternative to using 0.3 Jupiter mass to define a gas giant, as mass is missing in many instances of the NASA archives). Five features were selected as input predictors: metallicity [Fe/H] (st_met), temperature in Kelvin (st_teff), mass in solar units (st_mass), radius in solar units (st_rad), and surface gravity in logarithmic cgs units (st_logg). Two classification algorithms were trained and then evaluated against a dummy that baseline: logistic regression, which serves as an interpretable linear benchmark, and random forest, which captures nonlinear interactions among features. Interestingly, the random forest achieved the highest overall accuracy at 76.8% and the highest precision at 0.684, while the logistic regression excelled with recall (0.705) and F1 score (0.652). Feature importance analysis using both Gini impurity reduction and SHAP (SHapley Additive exPlanations) values identified surface gravity and radius as the two most influential predictors, with metallicity following close behind at third. These results are consistent with the core accretion theory of giant planet formation, which predicts that higher metallicity stars provide more solid building material for planet cores. At the same time, the results suggest that observational selection effects embedded in the archive also contribute notable predictive signals. These findings show that even a small set of stellar measurements, widely available to the public in catalogs and archives, can effectively and accurately distinguish likely gas giant hosts from non-hosts at accuracy levels well above those of chance, and that interpretable machine learning models can recover meaningful patterns from large datasets without requiring explicit astrophysical assumptions.