Your Cheat Sheet
The whole course on one page.
Here it is. Everything that you've learned, in the fewest possible words. Pause any of these screens, and take a screenshot. Keep them near your desk for your next project.
- The whole course, compressed
- Pause and screenshot
- Keep it near your desk
Which algorithm?
When you start a new problem, use this table. For regression on tables of data, begin with ridge as the baseline, then boosting. For classification, start with logistic regression, then boosting. For groups, try k means, then DBSCAN. For fewer dimensions, use P C A. And for strange cases, use isolation forest. Use the simplest thing that meets your need.
| Task | First try | Then |
|---|---|---|
| Regression | Ridge | Boosting |
| Classification | Logistic | Boosting |
| Grouping | K-Means | DBSCAN |
| Fewer dimensions | PCA | UMAP |
| Odd cases | Isolation forest | Autoencoder |
Which metric?
And here are the metrics. For regression, use M A E to explain the results, and R M S E when big errors hurt more. For balanced classes, accuracy is fine. For rare events, use precision and recall. For ranking risk, use the area under the curve. For money, use actual profit.
| Situation | Use |
|---|---|
| Regression | MAE, RMSE, R squared |
| Balanced classes | Accuracy, F1 |
| Rare events | Precision, recall, PR AUC |
| Ranking risk | ROC AUC |
| Business | Profit in dollars |
The recipe
And here's the recipe for every project. Define the decision. Split off the test set. Draw the data. Build a baseline. Put everything in a pipeline. Compare models with cross validation. Study your errors, and engineer features. Tune a little. Open the test set one time. Ship it, then monitor it, then retrain it.
Summary
That's the entire course. Screenshot those three screens, and you'll be ready for any new project. Next, let's clear up some myths that you'll hear about machine learning, and learn the jargon, so that you can hold your own in any meeting.
Key takeaways- Pick by the table
- Measure the right thing
- Follow the recipe