ENH: Random Forest Classifier oob scaling/parallel

Nobody has claimed this yet.

My team, working on a bioinformatics problem with high feature count (columns/dimensions in X), noticed that the RandomForestClassifier out of bag scoring doesn't scale with n_jobs. To be fair, n_jobs clearly says what it does support, though I do wonder if the out of bag predictions under the hood might also benefit from parallel support. Someone on my team seems to have found that it does help, but implemented externally to sklearn using the exposed base estimators. I suppose it might be nice to have that internally at some point, if there are no design reasons not to?

Sample reproducer code with latest stable release (1.3.2) on 16 cores/x86_64 Linux box (i9-13900K) is below the fold, and the scaling plot is underneath that. We also use far more estimators and features than that, so the delta is much greater, but the scaling trend is the main observation in any case.

from time import perf_counter
import numpy as np
# sklearn 1.3.2
from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import make_classification
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt


timings_oob = []
timings_base = []
feature_counts = np.linspace(10, 180_000, 20, dtype=np.int64)

for feature_count in feature_counts:
    X, y = make_classification(n_samples=1_000,
                               n_features=feature_count,
                               random_state=0)
    for use_oob, timing_list in zip([True, False], [timings_oob, timings_base]):
        start = perf_counter()
        clf = RandomForestClassifier(n_estimators=50,
                                     random_state=0,
                                     oob_score=use_oob,
                                     n_jobs=16)
        clf.fit(X, y)
        timing_list.append(perf_counter() - start)

fig, ax = plt.subplots(1, 1)
ax.set_title(f"Random Forest OOB scaling performance")
ax.plot(feature_counts,
        timings_oob,
        label="WITH OOB",
        marker=".")
ax.plot(feature_counts,
        timings_base,
        label="NO oob",
        marker=".")
ax.set_xlabel("num features")
ax.set_ylabel("Time (s)")
ax.legend()
fig.savefig("bench_feat.png", dpi=300)

Start from RandomForestClassifier.fit with oob_score=True and n_jobs set, using the issue's reproducer to compare OOB and non-OOB timing as feature counts grow. Read the existing OOB scoring path and its parallelization behavior; done means an agreed implementation or documented design decision, with tests or benchmarks showing the expected n_jobs scaling.

Written by the indexing model from the issue text.

Description

Contributor guide

First steps

Research direction

Assessment