Can't instanciate TreeExplainer when a tree in LightGBM Classifier model has n_features_in_tree = 1

[CharlesCousyn]

I have a LightGBM classifier working well.
But when i execute:

explainer  = shapiq.TreeExplainer(
    model=model.named_steps["classifier"], # My LightGBMClassifier
    max_order= 1,
    min_order= 1,
    index="k-SII",
    class_index=0
)

It fails with:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[6], line 1
----> 1 explainer  = shapiq.TreeExplainer(
      2     model=model.named_steps[\"classifier\"],
      3     max_order= 1,
      4     min_order= 1,
      5     index=\"k-SII\",
      6     class_index=0
      7 )
      8 shap_values, _ = explainer.explain(transformed_dataset)
     10 # Compute shap values
     11 # shap_values = feature_metrics.get_shap_values(model.named_steps[\"classifier\"], x, force_use_all_data=True)
     12 
     13 # Convert shap into readable dataframe

File ~/Code/isa-ml/isa-ml-jobs/.venv/lib/python3.10/site-packages/shapiq/explainer/tree/explainer.py:73, in TreeExplainer.__init__(self, model, max_order, min_order, index, class_index, **kwargs)
     70 self._class_label: Optional[int] = class_index
     72 # setup explainers for all trees
---> 73 self._treeshapiq_explainers: list[TreeSHAPIQ] = [
     74     TreeSHAPIQ(model=_tree, max_order=self._max_order, index=index) for _tree in self._trees
     75 ]
     76 self.baseline_value = self._compute_baseline_value()

File ~/Code/isa-ml/isa-ml-jobs/.venv/lib/python3.10/site-packages/shapiq/explainer/tree/explainer.py:74, in <listcomp>(.0)
     70 self._class_label: Optional[int] = class_index
     72 # setup explainers for all trees
     73 self._treeshapiq_explainers: list[TreeSHAPIQ] = [
---> 74     TreeSHAPIQ(model=_tree, max_order=self._max_order, index=index) for _tree in self._trees
     75 ]
     76 self.baseline_value = self._compute_baseline_value()

File ~/Code/isa-ml/isa-ml-jobs/.venv/lib/python3.10/site-packages/shapiq/explainer/tree/treeshapiq.py:122, in TreeSHAPIQ.__init__(self, model, max_order, min_order, index, verbose)
    120 if self._index in (\"SV\", \"SII\", \"k-SII\"):  # SP is of order at most d_max
    121     self.n_interpolation_size = min(self._edge_tree.max_depth, self._n_features_in_tree)
--> 122 self._init_summary_polynomials()
    124 # stores the nodes that are active in the tree for a given instance (new for each instance)
    125 self._activations: np.ndarray[bool] = np.zeros(self._n_nodes, dtype=bool)

File ~/Code/isa-ml/isa-ml-jobs/.venv/lib/python3.10/site-packages/shapiq/explainer/tree/treeshapiq.py:410, in TreeSHAPIQ._init_summary_polynomials(self)
    406 subset_ancestors: dict[int, np.ndarray] = self._precalculate_interaction_ancestors(
    407     interaction_order=order, n_features=self._n_features_in_tree
    408 )
    409 self.subset_ancestors_store[order] = subset_ancestors
--> 410 self.D_store[order] = np.polynomial.chebyshev.chebpts2(self.n_interpolation_size)
    411 self.D_powers_store[order] = self._cache(self.D_store[order])
    412 if self._index in (\"SV\", \"SII\", \"k-SII\"):

File ~/Code/isa-ml/isa-ml-jobs/.venv/lib/python3.10/site-packages/numpy/polynomial/chebyshev.py:1985, in chebpts2(npts)
   1983     raise ValueError(\"npts must be integer\")
   1984 if _npts < 2:
-> 1985     raise ValueError(\"npts must be >= 2\")
   1987 x = np.linspace(-np.pi, 0, _npts)
   1988 return np.cos(x)

ValueError: npts must be >= 2"
}

I identify that we get ValueError: npts must be >= 2 when one of the trees inside the model has the property n_features_in_tree = 1

[mmschlk]

Aaaah interesting! :/
This fails because of the chebyshev approximation of the summary polynomials. Apparently, this needs at least two features, or some 2d data in the numpy implementation. I will check this out! :)

[CharlesCousyn]

Due to the symmetry of Chebyshev nodes of the second kind (https://en.wikipedia.org/wiki/Chebyshev_nodes), does a value of 0 would be logical and coherent ?
With a change in the code in shapiq/explainer/tree/treeshapiq.py:410 looking like this?

...
if self.n_features_in_tree == 1:
    self.D_store[order] = np.array([0])  # Assign a default value
else:
    self.D_store[order] = np.polynomial.chebyshev.chebpts2(self.n_features_in_tree)
...

Do you think i will work to compute shap values?

[mmschlk]

That is a good idea. I will need to check this. another solution would be to simply don't do any computation because if tree has only one feature this feature gets all the contribution and there would not be any interaction or side effects. A try, except could work but would be ugly perhaps. :D