Epilogue

Seven chapters, seven models, and the basics of machine learning, one idea at a time.

Chapter 1 used logistic regression: a straight line drawn through a decision. Chapter 2 used a decision tree, a model you can read out loud, one split at a time. In Chapter 3, many trees voted together as a random forest, and in Chapter 4, gradient boosting added each new tree to fix what the last one got wrong. Chapter 5 set out to use a support vector machine, a model that looks for the widest possible gap between two outcomes, but the public data could not support the design it was chosen for. It answered two narrower questions with logistic regression instead. Chapter 6 used a neural network, a model that learns on its own which combinations of positions in a peptide matter. Chapter 7, on immune cells profiled by single-cell RNA-seq, used k-means: no answer key at all, only the data.

The first six were supervised. They learned from examples someone had already labelled. The last was unsupervised. It had to find the groups by itself, and it showed exactly where that gets hard. Sorting the cells, once it was told how many groups to make, was the easy part. Deciding how many groups there are was not.

The models changed from chapter to chapter. The approach did not. Fix the rules before fitting anything. Compare against a simple baseline. Share what did not work as openly as what did.

That approach is better than the model. A result nobody can check is only an opinion.

This epilogue is rewritten from the post for Day 70 of the series.