Leakage-Free Feature Standardization

Leakage-Free Feature Standardization

Standardization (z-score normalization) centers each feature to zero mean and unit variance:

The leakage trap: If you compute $\mu$ and $\sigma$ over both train and test data, test information contaminates your scaler — your validation metrics will be optimistic and the model will underperform in production.

Correct approach:

1. Fit on X_train only: compute per-column $\mu$ and $\sigma$ using np.mean(..., axis=0) and np.std(..., axis=0)

2. Transform X_test with the same $\mu$ and $\sigma$

Edge case: If any column has $\sigma =0$ (constant), set $\sigma =1$ to avoid division by zero.

Your task:

Implement scale_with_train_stats(X_train, X_test) that returns standardized X_test.