38. Overfitting and Underfitting in Machine Learning
38.3 Simple Fixes
Why. Once you name the problem, apply simple fixes before fancy theory.
What. Common fixes: more clean rows; fewer / better features; simpler model (shallower tree, fewer forest trees to start); regularisation options later; always keep a true holdout test set.
How.
Fix menu for beginners
- Overfit? Reduce tree depth; add data; drop leaky or noisy columns.
- Underfit? Add useful features; allow a slightly richer model; check you did not delete the signal.
- Always: evaluate on test rows you never trained on.
# Same API, simpler hypothesis
safer_tree = DecisionTreeClassifier(max_depth=3, min_samples_leaf=2, random_state=0)
Ravindra Bagale's Tip
Pehle simple model + clean split. Complexity baad me.