Derek Lowe, writing on models for protein-ligand interaction
Ideally, you train your model on a big pile o’ protein data, while holding out a good piece of it to keep it out of the training set. Then once your software has machine-learned its way to happiness, you see how it performs on those cases that it’s never seen before. But here’s the problem: it is very hard to set this up in a way so that the program is really seeing things that it hasn’t seen.
If you think about this from a statistical point of view, it seems strange. Splitting data into test and training sets at random automatically gives you independence, and we established nearly a century ago that random sampling is better than trying to engineer representativeness.
The problem is extrapolation. We aren’t trying to get good predictions for the distribution of problems where we know the answer, but for some larger universe of more interesting problems: all protein-ligand binding in humans or in eukaryotes or in living organisms. The problems where we know the ground truth tend to be both clustered and taken from a few parts of the problem space where the solution is easy.
To mimic, as closely as feasible, testing in an independent sample from the true universe of problems we don’t want to divide our ground-truth data up at random. We want to assign a whole cluster of examples to either training or test sets, and (ideally) to undersample where our ground-truth data has oversampled. This is kind of like the ideas about cross-validation for complex survey data, but more extreme. We don’t have sampling weights or sampling units identified in the data and we don’t know anything about the gaps in the data.
Since we can’t rely on random sampling to guarantee transportability from training to test to production data it’s also important that the model is ‘true for reasons’: that it captures correlations that come from causal mechanisms or that are inherited from common ancestors.