import pandas as pd import numpy as np import lightgbm as lgb import gc def haversine_distance(lat1, lon1, lat2, lon2): phi1, phi2 = np.radians(lat1), np.radians(lat2) dphi = np.radians(lat2 - lat1) dlambda = np.radians(lon2 - lon1) a = np.sin(dphi / 2)**2 + np.cos(phi1) * np.cos(phi2) * np.sin(dlambda / 2)**2 return (6371 * 2 * np.arcsin(np.sqrt(a))).astype('float32') def calculate_bearing(lat1, lon1, lat2, lon2): l1, l2 = np.radians(lat1), np.radians(lat2) diff_lon = np.radians(lon2 - lon1) x = np.sin(diff_lon) * np.cos(l2) y = np.cos(l1) * np.sin(l2) - (np.sin(l1) * np.cos(l2) * np.cos(diff_lon)) return np.degrees(np.arctan2(x, y)).astype('float32') def prepare_features(df): dt = pd.to_datetime(df['pickup_datetime'], format='%Y-%m-%d %H:%M:%S UTC', errors='coerce') df['hour'] = dt.dt.hour.fillna(0).astype('int8') df['dayofweek'] = dt.dt.dayofweek.fillna(0).astype('int8') df['year'] = (dt.dt.year.fillna(2012) - 2008).astype('int8') df['month'] = dt.dt.month.fillna(1).astype('int8') df['dayofyear'] = dt.dt.dayofyear.fillna(1).astype('int16') df['is_night'] = ((df['hour'] >= 20) | (df['hour'] <= 5)).astype('int8') df['is_rush'] = ((df['hour'] >= 16) & (df['hour'] <= 20) & (df['dayofweek'] < 5)).astype('int8') df['is_weekend'] = (df['dayofweek'] >= 5).astype('int8') df['fare_hike'] = (dt > '2012-09-04').astype('int8') del dt; gc.collect() # Clipping for valid NYC area df['pickup_latitude'] = df['pickup_latitude'].clip(40.5, 41.0) df['pickup_longitude'] = df['pickup_longitude'].clip(-74.3, -73.7) df['dropoff_latitude'] = df['dropoff_latitude'].clip(40.5, 41.0) df['dropoff_longitude'] = df['dropoff_longitude'].clip(-74.3, -73.7) df['lat_diff'] = (df['dropoff_latitude'] - df['pickup_latitude']).abs().astype('float32') df['lon_diff'] = (df['dropoff_longitude'] - df['pickup_longitude']).abs().astype('float32') df['haversine'] = haversine_distance(df['pickup_latitude'], df['pickup_longitude'], df['dropoff_latitude'], df['dropoff_longitude']) df['euclidean'] = np.sqrt(df['lat_diff']**2 + df['lon_diff']**2).astype('float32') df['bearing'] = calculate_bearing(df['pickup_latitude'], df['pickup_longitude'], df['dropoff_latitude'], df['dropoff_longitude']).fillna(0) # Rotated Manhattan Distance (NYC Grid is approx 29 degrees) rad = np.radians(28.92) p_lat_rot = df['pickup_latitude'] * np.cos(rad) - df['pickup_longitude'] * np.sin(rad) p_lon_rot = df['pickup_latitude'] * np.sin(rad) + df['pickup_longitude'] * np.cos(rad) d_lat_rot = df['dropoff_latitude'] * np.cos(rad) - df['dropoff_longitude'] * np.sin(rad) d_lon_rot = df['dropoff_latitude'] * np.sin(rad) + df['dropoff_longitude'] * np.cos(rad) df['rot_manhattan'] = ((p_lat_rot - d_lat_rot).abs() + (p_lon_rot - d_lon_rot).abs()).astype('float32') del p_lat_rot, p_lon_rot, d_lat_rot, d_lon_rot; gc.collect() # Landmark Distances landmarks = { 'jfk': (40.6413, -73.7781), 'lga': (40.7769, -73.8740), 'ewr': (40.6895, -74.1745), 'times_sq': (40.7580, -73.9855), 'gc': (40.7527, -73.9772), 'penn': (40.7505, -73.9934), 'port_auth': (40.7568, -73.9910), 'wall_st': (40.7060, -74.0088) } for name, coords in landmarks.items(): df[f'p_dist_{name}'] = haversine_distance(df['pickup_latitude'], df['pickup_longitude'], coords[0], coords[1]) df[f'd_dist_{name}'] = haversine_distance(df['dropoff_latitude'], df['dropoff_longitude'], coords[0], coords[1]) # Precise JFK Flat Fare Logic (Manhattan <-> JFK) p_man = (df['pickup_latitude'] > 40.70) & (df['pickup_latitude'] < 40.85) & (df['pickup_longitude'] > -74.02) & (df['pickup_longitude'] < -73.93) d_man = (df['dropoff_latitude'] > 40.70) & (df['dropoff_latitude'] < 40.85) & (df['dropoff_longitude'] > -74.02) & (df['dropoff_longitude'] < -73.93) df['jfk_flat_fare'] = ((p_man & (df['d_dist_jfk'] < 2.5)) | (d_man & (df['p_dist_jfk'] < 2.5))).astype('int8') # Interactions df['haversine_year'] = df['haversine'] * df['year'].astype('float32') df['haversine_hike'] = df['haversine'] * df['fare_hike'].astype('float32') df['haversine_month'] = df['haversine'] * df['month'].astype('float32') df['haversine_rush'] = df['haversine'] * df['is_rush'].astype('float32') df.drop(['pickup_datetime'], axis=1, inplace=True) return df def main(): dtypes = { 'fare_amount': 'float32', 'pickup_longitude': 'float32', 'pickup_latitude': 'float32', 'dropoff_longitude': 'float32', 'dropoff_latitude': 'float32', 'passenger_count': 'int8' } print("Loading data...") train = pd.read_csv('labels.csv', dtype=dtypes) train = train[(train['fare_amount'] >= 2.5) & (train['fare_amount'] <= 500)].reset_index(drop=True) y = train['fare_amount'].values train.drop('fare_amount', axis=1, inplace=True) test = pd.read_csv('test.csv', dtype={k: v for k, v in dtypes.items() if k != 'fare_amount'}) test_keys = test['key'].copy() combined = pd.concat([train, test], axis=0).reset_index(drop=True) n_train = len(train) del train, test; gc.collect() print("Preparing features...") combined = prepare_features(combined) if 'key' in combined.columns: combined.drop('key', axis=1, inplace=True) X = combined.iloc[:n_train] X_test = combined.iloc[n_train:] del combined; gc.collect() print("Training...") model = lgb.LGBMRegressor( n_estimators=2000, learning_rate=0.05, num_leaves=255, feature_fraction=0.75, subsample=0.75, reg_alpha=5.0, reg_lambda=5.0, n_jobs=8, verbose=-1, random_state=42 ) model.fit(X, y) del X, y; gc.collect() print("Predicting...") preds = np.clip(model.predict(X_test), 2.5, 500) sample_sub = pd.read_csv('sample_submission.csv') res_df = pd.DataFrame({'key': test_keys, 'fare_amount': preds}) submission = sample_sub[['key']].merge(res_df, on='key', how='left') submission['fare_amount'] = submission['fare_amount'].fillna(11.35) submission.to_csv('submission.csv', index=False) print("Done.") if __name__ == "__main__": main()

Taxi Fare RMSE

Step Code View