russell@unturf.com
25c2bafdee
New UNDF assignments (693→720): elixir-0002 → UNDF-2026-000000698 (typespec used_type_pairs O(T²)) r-source-0002 → UNDF-2026-000000711 (.walkClassGraph match dedup O(S²)) ruby-0003 → UNDF-2026-000000712 (RubyGems dependent_gems O(N²×D)) victoria-metrics-0002 → UNDF-2026-000000717 (MetricName tag-filter O(T×I)) Total: 720 UNDF assigned
67 lines
2.4 KiB
Markdown
67 lines
2.4 KiB
Markdown
# UNDF: UNDF-2026-000000715
|
||
# sklearn-0001: HistGradientBoosting _check_categories — O(n²) feature_names.index in loop
|
||
|
||
**Severity:** MEDIUM
|
||
**CWE:** CWE-407 (Algorithmic Complexity — linear membership test in hot loop)
|
||
**Speedup:** >40x at F=1000 features, C=200 categorical features
|
||
**Target:** scikit-learn (scikit-learn/scikit-learn)
|
||
**File:** `sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py:440-443`
|
||
|
||
## Description
|
||
|
||
`_check_categories` resolves categorical feature names to integer indices by
|
||
calling `feature_names.index(feature_name)` inside a `for feature_name in
|
||
categorical_features` loop. `feature_names` is a plain Python list built from
|
||
the DataFrame column names. Each call to `.index()` performs a linear scan from
|
||
position 0, making the total complexity O(C × F) where C is the number of
|
||
categorical features and F is the total feature count.
|
||
|
||
With wide feature-rich datasets (1000+ columns, 200+ categorical), this is called
|
||
at every `fit()`, `predict()`, and `score()` invocation — directly on the hot
|
||
path for model training and inference.
|
||
|
||
## Root Cause
|
||
|
||
```python
|
||
# sklearn/ensemble/_hist_gradient_boosting/gradient_boosting.py:440-443
|
||
is_categorical = np.zeros(n_features, dtype=bool)
|
||
feature_names = list(feature_names_in_) # plain list
|
||
for feature_name in categorical_features:
|
||
try:
|
||
is_categorical[feature_names.index(feature_name)] = True # O(F) per cat feature
|
||
except ValueError as e:
|
||
...
|
||
```
|
||
|
||
Fix: build a `dict` mapping name → index once before the loop.
|
||
|
||
## Patch
|
||
|
||
```python
|
||
is_categorical = np.zeros(n_features, dtype=bool)
|
||
feature_names = list(feature_names_in_)
|
||
feature_name_to_idx = {name: i for i, name in enumerate(feature_names)} # O(F) once
|
||
for feature_name in categorical_features:
|
||
try:
|
||
is_categorical[feature_name_to_idx[feature_name]] = True # O(1)
|
||
except KeyError:
|
||
raise ValueError(
|
||
f"categorical_features has an item value '{feature_name}' "
|
||
"which is not a valid feature name of the training "
|
||
f"data. Observed feature names: {feature_names}"
|
||
)
|
||
```
|
||
|
||
## Complexity Before
|
||
|
||
**O(C × F)** — C categorical features × F total features per index lookup
|
||
|
||
## Complexity After
|
||
|
||
Build index once: **O(F)**, then **O(1)** per lookup → **O(F + C)** total
|
||
|
||
## Reproduction
|
||
|
||
```
|
||
cd defects/sklearn/unit && javac -d . *.java && java -ea unit.SklearnTest
|
||
```
|