[ENH] Improve performance of isotonic regression

benchmarks/bench_isotonic.py

@@ -0,0 +1,31 @@
+from __future__ import print_function
+
+import numpy as np
+import gc
+from datetime import datetime
+from sklearn.isotonic import IsotonicRegression
+from sklearn.utils.bench import total_seconds
+import sys
+
+
+def bench_isotonic_regression(X, Y):
+    gc.collect()
+
+    tstart = datetime.now()
+    clf = IsotonicRegression()
+    clf.fit(X, Y)
+    delta = datetime.now() - tstart
+    return total_seconds(delta)
+
+if __name__ == '__main__':
+    min_exp, max_exp, iters = map(int, sys.argv[1:])
+    timings = []
+    for i in range(min_exp, max_exp):
+        n = 10 ** i
+        X = np.arange(n)
+        Y = np.random.randint(-50, 50, size=(n,)) \
+            + 50. * np.log(1 + np.arange(n))
+        times = [bench_isotonic_regression(X, Y) for i in range(iters)]
+        timing = (n, np.mean(times))
+        timings.append(timing)
+        print(n, np.mean(times))

doc/whats_new.rst

@@ -154,6 +154,10 @@ Changelog
      :func:`feature_selection.f_regression` when variates are not centered.
      By `VirgileFritsch`_.
 
+   - Significant performance improvements (more than 100x speedup for
+     large problems) in :class:`isotonic.IsotonicRegression` by
+     `Andrew Tulloch`_.
+
 
 API changes summary
 -------------------