adaptive-machine-learning / CapyMOA / 30998004410

80%

push

github

fix: stop discarding predictions with small vote totals

`MOAClassifier.predict_proba` returned None whenever the sum of
`getVotesForInstance` was <= 1e-2. MOA's votes are unnormalised and their
scale depends on the learner: majority-class leaves give integer counts,
while Naive Bayes gives products of likelihoods that sit far below that
threshold without being any less valid.

Measured over 2500 instances of SEA after a burn-in, NaiveBayes had a
median vote total of 6.29e-04, so the guard discarded 100% of its
predictions -- although the argmax of those same votes was 87.56% correct.
HoeffdingTree is affected through the Naive Bayes at its leaves: 80.2%
discarded with the default leaf prediction, 98.9% with NaiveBayes leaves,
and 0% with MajorityClass leaves.

Everything downstream saw None and treated it as no prediction. In
prequential evaluation that scored as a miss, so the Python loop and the
MOA loop disagreed by up to 52 accuracy points on the same stream and
learner, switched only by `optimise`.

The guard now rejects only votes that carry no prediction: an empty array,
a non-finite total, or nothing but zeros. The degenerate cases it was
introduced for are still rejected; small positive totals normalise as they
should.

Accuracy for affected learners improves because predictions that were
being thrown away are now counted. Pinned expectations are updated:
HoeffdingTree on TinySplitMNIST goes from 59.49 to 69.50 final accuracy,
and both loops now agree exactly on every stream tested.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

25 of 25 new or added lines in 2 files covered. (100.0%)

7864 of 10171 relevant lines covered (77.32%)

0.77 hits per line