From 1392d619bf05f97af040ab61063288f9f25b6d5b Mon Sep 17 00:00:00 2001
From: tommoral <thomas.moreau.2010@gmail.com>
Date: Thu, 9 Jul 2026 16:09:26 +0000
Subject: [PATCH] FIX CD plot: drop param-collapsing labels (rely on
 benchopt-shortened solver_name), declare scikit-posthocs requirement

---
 plots/plot_cd_grid.py            | 14 ++++----
 tests/plots/test_plot_cd_grid.py | 59 ++++++++++++++++++++++++++++++++
 2 files changed, 65 insertions(+), 8 deletions(-)
 create mode 100644 tests/plots/test_plot_cd_grid.py

diff --git a/plots/plot_cd_grid.py b/plots/plot_cd_grid.py
index 8c822d8..f9d9473 100644
--- a/plots/plot_cd_grid.py
+++ b/plots/plot_cd_grid.py
@@ -24,7 +24,6 @@
 import argparse
 import io
 import math
-import re
 import sys
 from pathlib import Path
 
@@ -42,10 +41,6 @@
 from benchmark_utils.metrics import is_higher_better  # noqa: E402
 
 
-def short_solver(name: str) -> str:
-    return re.sub(r"\[.*?\]$", "", str(name)).strip()
-
-
 def discover_metrics(df: pd.DataFrame) -> list[str]:
     """Return all numeric objective_* columns that have any non-NaN values."""
     cols = []
@@ -209,7 +204,6 @@ def _cd_global(df: pd.DataFrame, ax: plt.Axes, alpha: float = 0.05) -> str:
     k, n_blocks = stacked.shape
 
     mean_ranks = stacked.mean(axis=1)
-    mean_ranks.index = [short_solver(s) for s in mean_ranks.index]
 
     # Friedman on the stacked ranks. friedmanchisquare ranks within each
     # block; since we already supplied per-block ranks, the re-ranking is a
@@ -319,8 +313,7 @@ def cd_for_metric(
         [rankdata(rank_mat[c].values, method="average") for c in rank_mat.columns]
     ).T
     ranks_df = pd.DataFrame(ranks_arr, index=mat.index, columns=mat.columns)
-    mean_ranks = ranks_df.mean(axis=1).rename(short_solver)
-    mean_ranks.index = [short_solver(s) for s in mean_ranks.index]
+    mean_ranks = ranks_df.mean(axis=1)
 
     # Friedman test
     chi2, pval = friedmanchisquare(*[mat.iloc[i].values for i in range(k)])
@@ -498,6 +491,7 @@ class Plot(BasePlot):
 
     name = "Critical Difference Diagram"
     type = "image"
+    requirements = ["pip::scikit-posthocs"]
     options = {
         "objective_column": ...,
     }
@@ -522,6 +516,10 @@ def _get_all_plots(self, df):
         return plots, options
 
     def plot(self, df, objective_column):
+        # benchopt pre-shortens ``solver_name`` before ``plot`` is called
+        # (keeping only the parameters that vary), so several parametrizations
+        # of one solver arrive as distinct labels — the CD diagram groups on
+        # that column and gets unique short labels for free.
         fig, ax = plt.subplots(figsize=(10, 5))
         cd_for_metric(df, objective_column, ax)
         return [{"image": _fig_to_array(fig), "label": objective_column}]
diff --git a/tests/plots/test_plot_cd_grid.py b/tests/plots/test_plot_cd_grid.py
new file mode 100644
index 0000000..7145165
--- /dev/null
+++ b/tests/plots/test_plot_cd_grid.py
@@ -0,0 +1,59 @@
+"""Regression test for the CD-diagram plot.
+
+benchopt pre-shortens ``solver_name`` (keeping only the parameters that vary),
+so the plot receives unique labels and no longer collapses several
+parametrizations of one solver into a duplicate label — which used to make
+``scikit_posthocs`` raise "truth value of a Series is ambiguous".
+"""
+
+import sys
+from pathlib import Path
+
+import matplotlib
+import numpy as np
+import pandas as pd
+
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt  # noqa: E402
+
+BENCHMARK_DIR = Path(__file__).parents[2]
+sys.path.insert(0, str(BENCHMARK_DIR / "plots"))
+
+import plot_cd_grid as cd  # noqa: E402
+
+
+def _synthetic_results():
+    # Distinct solver names, as benchopt delivers them after short-labeling.
+    rng = np.random.default_rng(0)
+    solvers = [
+        "SeasonalNaive[season_length=1]",
+        "SeasonalNaive[season_length=7]",
+        "Chronos2",
+    ]
+    datasets = [f"d{i}" for i in range(6)]
+    return pd.DataFrame([
+        {
+            "solver_name": s,
+            "dataset_name": d,
+            "objective_mae": rng.random(),
+            "objective_mse": rng.random(),
+        }
+        for s in solvers
+        for d in datasets
+    ])
+
+
+def test_cd_for_metric_runs():
+    df = _synthetic_results()
+    fig, ax = plt.subplots()
+    status = cd.cd_for_metric(df, "objective_mae", ax)
+    plt.close(fig)
+    assert "objective_mae" in status
+
+
+def test_cd_global_runs():
+    df = _synthetic_results()
+    fig, ax = plt.subplots()
+    status = cd.cd_for_metric(df, cd.GLOBAL_KEY, ax)
+    plt.close(fig)
+    assert isinstance(status, str)