revanthkolla/demand-regression-lgbm

Demand Regression (LightGBM)

Gradient-boosted regression model predicting demand (continuous, range (0, 1]) from spatio-temporal + contextual tabular features. Trained on 77,299 rows.

Results (5-fold out-of-fold CV)

Metric	Value
R²	0.9608
RMSE	0.02815
MAE	0.01880

Best config (from a 4-point sweep): learning_rate=0.03, num_leaves=63, min_child_samples=30, reg_lambda=0.2, ~4309 trees. Target modeled as log1p(demand); predictions are expm1(...) clipped to ≥0.

Features

Raw columns: geohash, day, timestamp, RoadType, NumberofLanes, LargeVehicles, Landmarks, Temperature, Weather.

Engineered:

• geohash → lat/lon (decoded) + gh4/gh5 prefixes + raw geohash as categorical
• time-of-day: hour, minute, minute-of-day, 15-min slot, cyclical sin/cos
• missing flags (e.g. temp_missing); RoadType/Weather NaN → "missing" category
• LightGBM native categorical + native NaN handling for Temperature

Top features: geohash, Temperature, time-of-day (tod/sin/cos), lat/lon.

Usage

import pandas as pd
from huggingface_hub import hf_hub_download
# download predict.py from this repo, then:
from predict import load_model, predict

model = load_model()                 # downloads model.txt
df = pd.read_csv("new_data.csv")     # same raw columns as train (no 'demand')
df["demand_pred"] = predict(model, df)

Files

• model.txt — LightGBM booster
• predict.py — feature engineering + inference (handles geohash decode, NaNs)
• job_train.py / job_run.py — training pipeline (sweep → 5-fold OOF → final fit)
• metrics.json, sweep.json, feature_importance.json

Data

Trained from revanthkolla/ml-intern-1ebf8be8-datasets (config upload_b6c5285be689).