adaptive-machine-learning / CapyMOA / 38644732218

84%

push

github

fix(anomaly): stop StreamRHF from reseeding the global numpy generator

choose_split_attribute() called np.random.seed(node_seed) on every call -
once per tree node during forest construction and every window rebuild -
and the per-node split-value draw used np.random.uniform on the global
stream afterwards. Fitting a forest thus rewrote the random sequence of
anything else sharing the process (other learners, evaluation code, the
test harness itself), and identical detector inputs produced different
global RNG outcomes for unrelated code.

Use node-local np.random.RandomState instances seeded from the node's
seed, with the same uniform(0, Ks) -> uniform(min, max) draw order as
the legacy code, so tree structures and detector outputs are
bit-identical (verified by scoring a stream prefix on both sides of the
change). Add a contract test asserting that training leaves the global
numpy RNG state untouched.

Assisted-by: codebuff:deepseek-v4-flash

23 of 23 new or added lines in 2 files covered. (100.0%)

11290 of 13469 relevant lines covered (83.82%)

0.84 hits per line