Settings
Watch it: One real biological question, answered. Question 13 of instruction 59: which never-published proteins get a confident compartment call, from the question to the named genes. Silent screen capture of the real application.
Held-out labelcompartment
Error rate allowed (alpha)0.1
Base modellogistic
Guaranteeper class
Regularisation C0.5
Neighbours to consult15
The card strategy 35 opens on: the same four bars as every strategy (better than chance, reach, two plain task metrics, each with its interval and chance), Run / Test / Details ▸, what its test does, and one real failure beside one real success.
Console
strategy 35 test: Call genes with a stated error rate (split conformal prediction) -- started
conformal_calls: null 1 of 10
conformal_calls: null 2 of 10
conformal_calls: null 3 of 10
conformal_calls: null 4 of 10
conformal_calls: null 5 of 10
conformal_calls: null 6 of 10
conformal_calls: null 7 of 10
conformal_calls: null 8 of 10
conformal_calls: null 9 of 10
conformal_calls: null 10 of 10
conformal_calls: PASS
strategy 35 run: Call genes with a stated error rate (split conformal prediction) -- started
conformal_calls: done in 3.8s
PASS: set efficiency: 1 - (mean set size - 1) / (classes - 1) 0.757 against 0.251 ± 0.024 under 10 runs on shuffled labels, calibrated the same way (bar 0.288, p 0.0909); effect +0.506, needed +0.050; 951 hidden
Split conformal prediction at alpha 0.10: for at least 90% of new genes the prediction set contains the true compartment label. Of 4,336 unlabelled genes, 9 get a single label (called), 4,327 a set of several (the data cannot decide between them) and 0 an empty set (unlike any calibration gene).