Edit on GitHub

starplast.strategy_learning

Strategies 35-39: learners that say how sure they are, and learners that combine evidence.

35  Call genes with a stated error rate                (split conformal)
36  Smooth the measurements along the networks          (graph convolution + logistic)
37  Let a random forest find what defines a label        (random forest + permutation importance)
38  Learn how much to trust each kind of evidence        (stacked logistic regression)
39  Predict a value with an interval that holds          (gradient boosting + split conformal)

35 and 39 add what no other strategy offers: an error rate stated BEFORE the answer is seen. Conformal prediction turns any model's scores into sets (or intervals) that contain the truth for at least 1 - alpha of new genes, whatever the model, as long as the calibration genes resemble the new ones -- so a call is made only where the set is a single label, and the self-test checks the stated coverage on genes the calibration never saw. 36 lets a linear model see a gene's network neighbourhood; 37 lets a non-linear one find thresholds and interactions; 38 learns, out of fold, how far to trust each kind of evidence for this label.

Every calibration and every meta-model split is by whole orthogroups, as every test here is.

ADVANCED = 'Advanced models'

ALPHA = Param(name='alpha', kind='float', label='Error rate allowed (alpha)', tip='The share of new genes whose true answer may fall outside what is returned. 0.1 promises that at least 90% of prediction sets contain the true label (or 90% of intervals the true value); smaller alpha gives bigger sets and wider intervals.', default=0.1, lo=0.01, hi=0.5, step=0.01, choices=(), optional=False, space='this')