Stargazer AI — Dataset & Training Guide

← Back to Home

Overview — ML pipeline (brief)

1. Upload a labelled dataset (CSV) that contains the target column and features.
2. Preprocess: handle missing values, convert types, encode categorical columns, scale numeric features.
3. Feature engineering: create derived features the model benefits from (example: ratios, log transforms).
4. Balance classes if needed (SMOTE is used here) and run cross-validation.
5. Train candidate models, evaluate with cross-validation, select the best model and inspect metrics and confusion matrix.
6. Upload unlabelled test data and run predictions using the same preprocessing pipeline.

Required / recommended features

The pipeline expects features similar to the list below. Not every dataset will use the exact same names — mapping/renaming columns is common.

• koi_disposition (target: e.g. CANDIDATE, FALSE POSITIVE, CONFIRMED)
• koi_period
• koi_time0bk
• koi_impact
• koi_duration
• koi_depth
• koi_ror
• koi_srho
• koi_prad
• koi_sma
• koi_incl
• koi_num_transits
• koi_steff
• koi_slogg
• koi_smet
• koi_srad
• koi_smass
• koi_kepmag, koi_gmag, koi_rmag, koi_imag, koi_zmag
• koi_jmag, koi_hmag, koi_kmag

Column-name differences

If your dataset uses different column names, rename them to the ones above before uploading. Common examples:

• period → koi_period
• depth → koi_depth
• duration → koi_duration
• prad → koi_prad

Cleaning & preprocessing checklist

• Remove duplicate rows and obvious corrupt records.
• Ensure the target column (koi_disposition) exists and has consistent labels.
• Numeric columns: impute missing values (median recommended), clip extreme outliers (IQR rule) or winsorize.
• Categorical columns: fill missing with the mode or the token 'UNKNOWN', then encode (LabelEncoder or one-hot).
• Scale numeric features (StandardScaler) before training models that expect scaled inputs.
• Create simple derived features used by the app: planets_to_star_radius_ratio = koi_prad / koi_srad, log_period = log1p(koi_period), depth_to_duration = koi_depth / koi_duration.
• Split a hold-out test set or use cross-validation (the app uses stratified CV + SMOTE for class imbalance).

Quick data-mapping example

If your CSV has different names, you can prepare a small mapping script before upload (example using pandas):

import pandas as pd
df = pd.read_csv('my_data.csv')
rename_map = {'period':'koi_period', 'depth':'koi_depth', 'duration':'koi_duration', 'prad':'koi_prad'}
df.rename(columns=rename_map, inplace=True)
df.to_csv('my_data_renamed.csv', index=False)
            

Training tips

• Keep a small validation split or use stratified folds to preserve class distribution.
• Try these metrics: accuracy, precision, recall, F1 (macro) and review the confusion matrix.
• Use class balancing (SMOTE) if classes are imbalanced; monitor for overfitting.
• For the XGBoost model used in this app: tune n_estimators, max_depth, and learning_rate.

Recommended CSV formatting

• Comma-separated, UTF-8, with a header row.
• Use consistent missing value markers (empty cells are fine).
• Include the target column for labelled uploads.

If you want, I can add a small helper page to help rename and preview a mapping interactively before upload — would you like that?