Model Evaluation & Metrics
Debugging ML Models
5 min read
Common Issues
High train, low test accuracy? → Overfitting
- Solutions: Regularization, more data, simpler model, dropout
Low train, low test accuracy? → Underfitting
- Solutions: More complex model, more features, less regularization
Perfect accuracy on validation? → Data leakage
- Check: features from future, test data in training, target encoding before split
High variance? → Model too complex or small dataset High bias? → Model too simple
Interview Q: "Test acc 99%, train acc 60%. What's wrong?" A: Data leakage. Test performance shouldn't exceed training.
:::