Model Training

AirSense ML trains 7 models in a single pipeline run and automatically selects the best performer.

---

Running Training

make train

Trains all 7 models, logs everything to MLflow, saves the best model to models/best_model.pkl.

For production-optimised training (XGBoost + LightGBM only):

make train-prod

---

Models

Two families are trained — linear models as interpretable baselines, tree ensembles for performance.

Linear Family

Model	Purpose	Key Parameter
Linear Regression	Raw baseline — no regularisation	None
Ridge	L2 regularisation — handles multicollinearity	alpha
Lasso	L1 regularisation — automatic feature selection	alpha
ElasticNet	Combined L1+L2	alpha, l1_ratio

Linear models establish the performance floor. If tree models don't significantly outperform linear models, it suggests the relationship is mostly linear and simpler models are sufficient.

Tree Ensemble Family

Model	Strength	Memory
Random Forest	Stable, parallel trees, interpretable importance	~400MB
XGBoost	Sequential boosting, speed + accuracy	~50MB
LightGBM	Fastest, leaf-wise growth, best for large data	~15MB

Tree models handle non-linear interactions automatically — they learn "high humidity + low wind + winter = severe AQI" without manual feature engineering.

---

Why Tree Models Outperform Linear by 45%

AQI is not a linear function of weather. The relationship depends on combinations:

• High humidity + low wind speed + winter → severe smog
• High temperature + monsoon season → pollution washout
• Night hours + residential areas → different baseline than daytime

Linear models can only learn AQI = a*temperature + b*humidity + .... They cannot learn "if temperature AND humidity AND season interact in this specific way". Tree models split on combinations of features, learning exactly these patterns.

Baseline results:

Model	Val RMSE	R²
Linear Regression	41.28	0.942
Random Forest	22.23	0.983

Random Forest reduces error by 46% over linear regression.

---

MLflow Experiment Tracking

Every training run logs:

• Parameters — model hyperparameters from config
• Metrics — val_rmse, val_r2, val_mae, val_rmsle, test_rmse, test_r2, cv_rmse
• Artifacts — the full serialised pipeline (features + model)

View all runs:

make mlflow
# Open http://127.0.0.1:5000

Experiments are stored in mlruns.db (SQLite). Never commit this file — it's gitignored.

---

Model Selection

The best model is selected by lowest validation RMSE — not test RMSE.

Test set is held out completely and used only for final reporting. Using test RMSE for selection would leak test information into the model choice.

# Primary metric from model_config.yaml
primary_metric: rmse

# Selection logic in train.py
if val_metrics.rmse < best_metric:
    best_metric = val_metrics.rmse
    best_model_name = model_name
    best_pipeline = full_pipeline

---

Cross-Validation

5-fold cross-validation runs on the training set for each model. This gives a more reliable estimate of generalisation than a single train/val split.

CV RMSE is reported in log space (the model is trained on log1p(AQI)) — it cannot be directly compared to Val/Test RMSE which is reported on the original AQI scale.

Use CV RMSE only to compare model consistency across folds — lower CV RMSE means more stable performance.

---

Serialisation

The full sklearn Pipeline object (feature transformer + model) is saved with joblib:

models/best_model.pkl

The pipeline contains everything needed for inference: - ColumnTransformer with fitted StandardScaler and OneHotEncoder - Fitted model (Random Forest, XGBoost, or LightGBM)

This guarantees that inference applies identical transformations to training — no training/serving skew.

---

Configuration

All training parameters are in configs/model_config.yaml:

training:
  test_size: 0.2
  val_size: 0.1
  random_state: 42
  cv_folds: 5
  shuffle: true

models:
  random_forest:
    enabled: true
    params:
      n_estimators: 200
      max_depth: 15
      ...

evaluation:
  primary_metric: rmse

mlflow:
  experiment_name: airsense-aqi-regression
  tracking_uri: sqlite:///mlruns.db

artifacts:
  model_dir: models/
  best_model_name: best_model.pkl

To disable a model, set enabled: false. To change hyperparameters, update params. No code changes required.