Classifying Potentially Hazardous Asteroids from NASA JPL Orbital Data Machine Learning Term Project — Group 16

Problem Statement NASA's Jet Propulsion Laboratory designates a subset of catalogued asteroids as Potentially Hazardous Asteroids (PHAs) — fewer than 1% of objects in the Small-Body Database. This project investigates whether the PHA designation can be predicted from orbital and physical parameters using supervised machine learning, and whether that predictive signal survives once the parameters that define the PHA criterion (minimum orbit intersection distance and absolute magnitude) are withheld. Dataset Source: NASA JPL Small-Body Database asteroid catalogue, via Kaggle Setup: place the CSV at the path set by DATA_PATH in the config cell (defaults to dataset.csv in the working directory) Target: pha (Y/N), encoded as binary (1 = hazardous) Environment Python 3 with: numpy, pandas, matplotlib scikit-learn imbalanced-learn (for the SMOTE comparison in Section 7.2) Key configuration variables (top of the notebook): RANDOM_STATE = 42 — fixed seed for reproducibility SAMPLE_ROWS = 150_000 — stratified subsample size used for the full workflow (set to None to use the complete catalogue) TUNE_ROWS = 30_000 — stratified slice used for cross-validation and hyperparameter search Methodology Data cleaning — drop identifier/metadata columns and duplicate date fields, resolve and encode the target, drop unlabeled records, assess field completeness. EDA — class distribution and imbalance, distribution of key orbital parameters by class, hazard prevalence by dynamical class, collinearity among orbital elements. Preprocessing — stratified 80/20 train/test split; features grouped into plain numeric, skewed numeric (signed-log + standardize), physical measurements (impute + missingness indicator), and categorical (one-hot); all steps wrapped in scikit-learn Pipeline/ColumnTransformer objects fit only on training data. Modeling — three classifiers compared: Logistic Regression, Random Forest, and Histogram-based Gradient Boosting, with class imbalance handled via balanced class weighting (SMOTE evaluated as an alternative). Hyperparameters tuned with RandomizedSearchCV under stratified 5-fold cross-validation, scored on average precision (PR AUC). Three experiments, evaluated once on the held-out test set: A — Complete feature set B — Excluding moid, moid_ld, H (the parameters that define the PHA rule) C — Orbital elements only (B, plus removing orbit_class and neo, which encode Earth-proximity indirectly) Key Results Experiment Best model PR AUC (average precision) A: all features Histogram Gradient Boosting 0.9949 B: excluding moid & H Random Forest 0.3846 C: orbital elements only Random Forest 0.4073 Near-perfect performance in Experiment A is definitional: the models recover the JPL threshold rule on moid and H rather than learning a genuinely empirical relationship. Performance drops substantially once those fields are withheld, but a real signal remains in the raw orbital elements (semi-major axis, eccentricity, perihelion distance, etc.). Ensemble methods (Random Forest, Hist Gradient Boosting) retain far more performance than Logistic Regression as features are removed, indicating the residual relationship is non-linear. Histogram Gradient Boosting is the recommended model overall — best on the full feature set and competitive with Random Forest on the reduced sets, at a fraction of the training time. See Section 12 of the notebook for full limitations (observational selection bias in the catalogue, sparse physical measurements, small positive test-set count, single train/test split, threshold left at default 0.5, etc.) and suggested further work (recall-targeted threshold selection for planetary-defense use cases). Notebook Structure Introduction / problem statement Environment and configuration Data acquisition and quality assessment Exploratory data analysis Methodology overview Outlier analysis, partitioning and preprocessing Model implementation, class imbalance and hyperparameter tuning Experiment A — evaluation on the complete feature set Feature importance and Experiment B Experiment C — orbital elements in isolation Comparative results Conclusions and limitations Contribution summary How to Run Download the dataset CSV from Kaggle (link above) and place it at the path set by DATA_PATH, or update DATA_PATH to point to your copy. Install dependencies:

   pip install numpy pandas matplotlib scikit-learn imbalanced-learn

Run all cells top to bottom in Jupyter (each data-processing cell assigns a new dataframe name rather than reassigning, so the notebook is largely idempotent under re-execution). Contributions Author Sections Focus Leonardo Rodrigues Fabbro 3, 12 Data loading/schema checks, data dictionary, cleaning, target encoding, duplicate/completeness checks, subsampling, conclusions Madhurishitha Boddu 4, 8 EDA (class distribution, feature distributions, dynamical class, correlation), evaluation function, test-set metrics, confusion matrices, PR curves, comparative results Shunmuka Valsa 6, 10 Outlier analysis, partitioning, feature grouping, signed-log transform, imputation/missingness indicators, one-hot encoding, ColumnTransformer, Experiment C Kunal Sharma 7, 9 Model pipelines, class weighting, SMOTE evaluation, cross-validation, hyperparameter search, refitting, feature importance, Experiment B All members contributed to the proposal, attended project meetings, and participated in the presentation.