Summary
We’ve spent some time looking at how to improve complex classification models, both with balanced data and with imbalanced data. We've learned that we identify issues and improve our models by:
- Better assessing the kinds of mistakes model is making.
- Rebalancing our data or altering the way our model is assessed.
- Changing the model architecture.
- Working with hyperparameters.
When working with complex data, machine learning experts often dedicate the most time to altering model architecture and working with hyperparameters to improve their models. We’ve seen how the wrong settings can hurt or improve model performance. A major factor deciding this is the size of the dataset in question. Often when we have smaller datasets, tuning architecture and hyperparameters can make sizeable improvements to models. With very large datasets, these can still often squeeze a small amount of performance gain out of our models.