Fix radar_plot axis ranges for asymmetric padding and constant metrics

[tonycoder-hub]

Problem

radar_plot derives each variable's axis range from the observed min/max plus 10% padding:

min_max_per_variable["min"] = min_max_per_variable["min"] - 0.1 * (
    min_max_per_variable["max"] - min_max_per_variable["min"]
)
min_max_per_variable["max"] = min_max_per_variable["max"] + 0.1 * (
    min_max_per_variable["max"] - min_max_per_variable["min"]
)

Two things go wrong here:

1. Asymmetric padding. The second statement recomputes the span using the min that the first statement already shifted downwards, so the top of every axis is padded by 11% of the data range while the bottom is padded by 10%. For accuracy values of 0.0 and 10.0 the axis ends up as (-1.0, 11.1) instead of (-1.0, 11.0).

2. Zero-width range for constant metrics. When a metric has the same value for every model the span is 0, so the padding is 0 too and the range collapses to a single point. Matplotlib then warns Attempting to set identical low and high ylims makes transformation singular, and because ComplexRadar._scale_data rescales all other variables into ranges[0], a constant first variable flattens every model onto the same radius — two models with different scores are drawn as identical shapes. If every metric is constant (e.g. a single model), _scale_data divides by zero and the plotted coordinates become NaN.

Fix

Compute the padding once from the original min/max, and fall back to a non-zero padding for variables that only take a single value.

Tests

Two tests added to tests/test_viz.py, both failing before the change and passing after:

• test_range_padding_is_symmetric asserts the accuracy axis is (-1.0, 11.0).
• test_metric_with_identical_values asserts that when accuracy is identical across models but precision is not, the two plotted shapes are finite and still differ.

$ MPLBACKEND=Agg python -m pytest tests/test_viz.py -q
4 passed

black, isort and flake8 (repo settings, line length 119) are clean on both changed files.

  

[tonycoder-hub]

Closing as stale — opened on or before 2026-08-17 and still unmerged.