Fix DiceGenetic.compute_proximity_loss for all-categorical datasets (#276)

[jbbqqf]

Summary

DiceGenetic.compute_proximity_loss divides by sum(feature_weights),
where feature_weights is restricted to the continuous feature indexes.
For an all-categorical dataset (e.g. survey responses, OHE-only inputs)
that array is empty — and the original code hit either of two bugs
depending on input shape:

• proximity_loss / sum(feature_weights) ⇒ ZeroDivisionError /
  RuntimeWarning + NaN (the snippet @kburchfiel quoted in proximity loss when there are not continous features #276), or
• product.reshape(-1, product.shape[-1]) ⇒ ValueError on the 0-sized
  array.

Both poison compute_loss with NaN/exceptions and break genetic search
for legitimate categorical-only use.

Why

Proximity is conceptually undefined when there are no continuous distances
to weigh, so the function short-circuits with a zero loss vector matching
the population shape. The categorical_penalty term in compute_loss
already handles categorical sparsity, so dropping the proximity
contribution is the correct semantic — and matches what users expect when
they explicitly set up a categorical-only dice_ml.Data.

The change is one early-return on len(feature_weights) == 0. Comment in
the code explains the choice for future reviewers.

Reproduce BEFORE/AFTER yourself (copy-paste)

set -e
cd /tmp && rm -rf DiCE-276 && git clone https://github.com/interpretml/DiCE.git DiCE-276
cd DiCE-276
pip install -q -e . pytest scikit-learn

git fetch https://github.com/jbbqqf/DiCE.git fix/276-proximity-loss-no-continuous
git checkout FETCH_HEAD -- tests/test_dice_interface/test_dice_genetic.py

# --- BEFORE: production code from origin/main ---
git checkout origin/main -- dice_ml/explainer_interfaces/dice_genetic.py
python -m pytest \"tests/test_dice_interface/test_dice_genetic.py::TestComputeProximityLossNoContinuousFeatures::test_compute_proximity_loss_returns_zero_when_no_continuous_features\" -q || echo \"BEFORE: ValueError 'cannot reshape array of size 0' (expected)\"

# --- AFTER: fix applied ---
git checkout FETCH_HEAD -- dice_ml/explainer_interfaces/dice_genetic.py
python -m pytest \"tests/test_dice_interface/test_dice_genetic.py::TestComputeProximityLossNoContinuousFeatures::test_compute_proximity_loss_returns_zero_when_no_continuous_features\" -q
# Expected: 1 passed

What I ran locally

• New test passes on this branch
• Test fails on origin/main with ValueError: cannot reshape array of size 0 into shape (0) (the original snippet's proximity_loss / sum(...) path is reached only with a slightly different input shape, but both have the same root cause)
• Full tests/test_dice_interface/test_dice_genetic.py ran: same 4 pre-existing failures as origin/main (regression suite for TestDiceGeneticRegressionMethods is broken on main for separate reasons), 0 new failures introduced

Edge cases

dataset shape	path taken	result
Mixed continuous + categorical	unchanged	unchanged
All categorical (this fix)	new early-return	np.zeros(population)
All continuous	unchanged	unchanged
proximity_weight = 0	already short-circuited in compute_loss	unchanged

AI disclosure

This change was prepared with the assistance of Claude (Anthropic).
The author reviewed every line and is responsible for the final result.