Speed up DBSTREAM by removing deepcopy from cleanup and reclustering

[MaxHalford]

Summary

• cProfile of the synthetic_sklearn workload showed copy.deepcopy(self._micro_clusters) inside _cleanup accounting for ~91% of learn_one time, with a second copy.deepcopy in _generate_clusters_from_labels dominating the predict path.
• This PR replaces both deepcopies, hoists the Gaussian neighborhood factor out of the per-feature center update loop (it's a single scalar, not per-dimension), and folds the nested try/except KeyError shared-density update into a plain dict.get / lazy-init.
• Output is unchanged: all 10 test_dbstream.py tests pass, including test_dbstream_synthetic_sklearn (v_beta checked to 4 decimal places).
• On the 15k-sample synthetic_sklearn workload, learn_one is ~6.1× faster (0.516 s → 0.084 s) and learn_one + predict_one is ~4.3× faster (0.872 s → 0.204 s).

Profile evidence (before)

   ncalls  tottime  cumtime  filename:lineno(function)
    15000    0.046    2.201  river/cluster/dbstream.py:253(_cleanup)
1920316/15000 0.981   2.147  copy.py:119(deepcopy)

_cleanup was 91% of total learn_one time, entirely from copy.deepcopy(self._micro_clusters) cloning every micro cluster on every cleanup tick just so a few could be popped. The fix is to collect weak ids in a first pass, then pop in place.

A second copy.deepcopy lived inside _generate_clusters_from_labels, called from every _recluster on the predict path. It was used purely to obtain a writable copy of the first micro cluster before merging the rest into it; building a fresh DBSTREAMMicroCluster with a shallow-copied center dict has the same semantics for the same price as a plain dict copy. While there, the O(k·m) double loop over range(max_label + 1) × cluster_labels.items() is collapsed to a single grouping pass.

Changes

• _cleanup: collect weak ids in one pass, then pop in place. No deepcopy.
• _generate_clusters_from_labels: group by label in one pass; build the macro cluster as a fresh DBSTREAMMicroCluster with a shallow-copied center dict.
• _update: hoist _gaussian_neighborhood(x, center) out of the per-feature center dict comprehension (it returned the same scalar for every j and was being recomputed n_features times); replace the nested try/except KeyError shared-density update with a dict.get / lazy-init.
• Drops the now-unused import copy.

Test plan

• uv run pytest river/cluster/test_dbstream.py river/cluster/dbstream.py — 10/10 pass (incl. doctest and the synthetic_sklearn v_beta check to 4 decimals)
• cProfile + timing harness on 15k samples (see Summary) — learn_one ~6.1× faster, learn_one + predict_one ~4.3× faster
• CI

🤖 Generated with Claude Code