Two simple plots answer common questions about model development.
Learning Curves
Plot training and validation performance as training set size grows.
- Both curves converge at poor performance: high bias. More data won't help much; use a more expressive model or better features.
- Large gap, validation still improving: high variance. More data is likely to help, as is regularisation.
- Curves converge at good performance: the model is in good shape.
Validation Curves
Plot training and validation performance as one hyperparameter changes — tree depth, regularisation strength, number of neighbours.
- Rising training score with falling validation score shows overfitting.
- The validation peak suggests a good setting.
Why They're Useful
They help decide whether to spend effort on collecting data, engineering features or tuning — before spending it.
Practical Tips
- Use cross-validation for smoother curves.
- Show variability with shaded bands.
- Libraries such as scikit-learn provide helpers to compute these curves.