sheshe 0.1.7

SheShe: Smart High-dimensional Edge Segmentation & Hyperboundary Explorer

SheShe

Smart High-dimensional Edge Segmentation & Hyperboundary Explorer

SheShe transforms any probabilistic model into a guided explorer of its own decision landscape. By following the local maxima of the class probability (classification) or the predicted value (regression), it discovers crisp, human‑readable regions that obey the supervised boundary of the problem. Rather than grouping samples by raw feature distance, SheShe learns from labeled data and carves clusters directly on top of the model’s decision surface.

Highlights

• Supervised clustering driven by the model’s own probabilities or predictions.
• Unified support for classification and regression tasks.
• Subspace exploration with SubspaceScout and ensembles via ModalScoutEnsemble.
• Human-readable rule extraction through RegionInterpreter.
• Built-in plotting utilities for pairwise and 3D visualisations.

Feature overview figure omitted (binary assets are not allowed).

---

Installation

Requires Python >=3.9 and it is recommended to work inside a virtual environment. Install the latest release from PyPI:

pip install sheshe

Base dependencies: numpy, pandas, scikit-learn>=1.1, matplotlib

For a development environment with tests:

pip install -e ".[dev]"
PYTHONPATH=src pytest -q

Reproducibility

This project was developed and tested on:

• OS: Ubuntu 24.04.2 LTS
• CPU: Intel(R) Xeon(R) Platinum 8171M CPU @ 2.60GHz
• GPU: None
• Python: 3.12.10

To recreate the environment:

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
PYTHONPATH=src pytest -q

---

Quick API

The library exposes five main objects:

• ModalBoundaryClustering
• ClusterRegion – dataclass describing a discovered region
• SubspaceScout
• ModalScoutEnsemble
• RegionInterpreter – turn ClusterRegion objects into human-readable rules

Figures illustrating these objects are omitted because binary assets are not allowed in this repository.

from sheshe import (
    ModalBoundaryClustering,
    SubspaceScout,
    ModalScoutEnsemble,
    ClusterRegion,
    RegionInterpreter,
)

# classification
clf = ModalBoundaryClustering(
    base_estimator=None,           # default LogisticRegression
    task="classification",         # "classification" | "regression"
    base_2d_rays=24,
    direction="center_out",        # "center_out" | "outside_in"
    scan_radius_factor=3.0,
    scan_steps=24,
    smooth_window=None,             # optional moving average window
    drop_fraction=0.5,              # fallback drop from peak value
    stop_criteria="inflexion",     # or "percentile" for percentile-bin drop
    percentile_bins=20,             # number of percentile bins when stop_criteria="percentile"
    random_state=0
)

# regression (example)
reg = ModalBoundaryClustering(task="regression")

Methods

• fit(X, y)
• predict(X)
• fit_predict(X, y=None) → convenience method equivalent to calling fit followed by predict on the same data
• predict_proba(X) → classification: per-class probabilities; regression: normalized value [0,1]
• decision_function(X) → decision scores from the base estimator; falls back to predict_proba for classification or predict for regression
• interpretability_summary(feature_names=None) → DataFrame with:
  • Type: "centroid" | "inflection_point"
  • Distance: radius from the center to the inflection point
  • Category: class (or "NA" in regression)
  • slope: df/dt at the inflection point
  • real_value / norm_value
• coord_0..coord_{d-1} or feature names
• predict_regions(X, label_path=None) → cluster ID(s) for each sample
• get_cluster(cluster_id) → retrieve a stored ClusterRegion
• plot_pairs(X, y=None, max_pairs=None) → 2D plots for all pair combinations
• save(filepath) → save the model using joblib
• ModalBoundaryClustering.load(filepath) → load a saved instance

Example of fit_predict usage:

from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering

X, y = load_iris(return_X_y=True)
labels = ModalBoundaryClustering().fit_predict(X, y)
print(labels[:5])

Regression example with retraining

from sklearn.datasets import load_diabetes
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestRegressor
from sheshe import ModalBoundaryClustering

X, y = load_diabetes(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0)

# initial training with the default estimator
reg = ModalBoundaryClustering(task="regression").fit(X_train, y_train)
print(reg.predict(X_test)[:3])

# retrain using a different base estimator
reg_retrained = ModalBoundaryClustering(
    base_estimator=RandomForestRegressor(random_state=0),
    task="regression",
).fit(X_train, y_train)
print(reg_retrained.predict(X_test)[:3])

decision_function(X)

Returns decision values from the underlying estimator. For classification it prefers the estimator's decision_function but falls back to predict_proba when that method is missing. In regression the method relies on predict as a fallback.

from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering

X, y = load_iris(return_X_y=True)
sh = ModalBoundaryClustering().fit(X, y)
print(sh.decision_function(X[:5]))

predict_regions(X, label_path=None)

Return cluster identifiers for each sample based solely on the discovered regions.

from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering

X, y = load_iris(return_X_y=True)
sh = ModalBoundaryClustering().fit(X, y)
print(sh.predict_regions(X[:3]))

get_cluster(cluster_id)

Fetch a stored :class:ClusterRegion by its identifier.

reg = sh.get_cluster(0)
print(reg.center)

Per-cluster metrics

After fitting, ModalBoundaryClustering stores the discovered regions in the regions_ attribute. Each ClusterRegion includes:

• score: effectiveness of the estimator on samples inside the region (accuracy for classification, R² for regression)
• metrics: optional dictionary with additional per-cluster metrics such as precision, recall, F1, MSE or MAE

Interpretability

RegionInterpreter – interpret cluster regions

from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering, RegionInterpreter

iris = load_iris()
X, y = iris.data, iris.target

sh = ModalBoundaryClustering().fit(X, y)
cards = RegionInterpreter(feature_names=iris.feature_names).summarize(sh.regions_)
RegionInterpreter.pretty_print(cards[:1])

Each card includes a cluster_id to identify the region and the class label.

OpenAIRegionInterpreter – describe regions with LLMs

Install the optional openai dependency (version >=1) and provide an API key using the api_key argument or via environment variables. The interpreter looks for OPENAI_API_KEY or OPENAI_KEY and, when running on Google Colab, also checks google.colab.userdata. Language and temperature defaults can be configured on the interpreter and overridden at call time. The layout parameter lets you enforce a general output template (for example, "bullet list") or omit it for free‑form text. Then call describe_cards to obtain natural‑ language explanations for the region cards.

from sheshe import OpenAIRegionInterpreter

expl = OpenAIRegionInterpreter(model="gpt-4o-mini", language="en", temperature=0.2)
texts = expl.describe_cards(cards, layout="bullet list", temperature=0.5)
print(texts[0])

Visualización 3D

plot_pair_3d visualiza la probabilidad de una clase o el valor predicho como una superficie tridimensional para un par de características.

Parámetros principales:

• pair: tupla (i, j) con los índices de las características a graficar.
• class_label: etiqueta de la clase a mostrar cuando task='classification'.
• grid_res: resolución de la malla usada para la superficie.
• alpha_surface: transparencia de la superficie.
• engine: 'matplotlib' (por defecto) para una figura estática o 'plotly' para un gráfico interactivo.

Ejemplo mínimo:

import matplotlib.pyplot as plt
from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering

iris = load_iris()
X, y = iris.data, iris.target

sh = ModalBoundaryClustering().fit(X, y)
# Modo estático con Matplotlib
sh.plot_pair_3d(X, (0, 1), class_label=sh.classes_[0])
plt.show()

# Modo interactivo con Plotly
fig = sh.plot_pair_3d(X, (0, 1), class_label=sh.classes_[0], engine="plotly")
fig.show()

---

How does it work?

1. Train/use a base model from sklearn (classification with predict_proba or regression with predict).
2. Find local maxima via gradient ascent with barriers at the domain boundaries.
3. From the maximum, trace rays (directions) on the hypersphere:
   • 2D: 24 rays by default
   • 3D: ~26 directions (coverage by spherical caps using Fibonacci sampling)

   • 3D: mixture of a few global directions + 2D/3D subspaces

4. Along each ray, scan radially and compute the first inflection point according to direction and stop_criteria:
   • center_out: from the center outward
   • outside_in: from the outside toward the center Optionally apply a moving average (smooth_window) and record the slope (df/dt) at that point. With stop_criteria="percentile" the scan stops when the value falls to a lower percentile bin of the dataset distribution (20 bins by default). If no stop is found, use the first point where the value drops below drop_fraction of the peak.
5. Connect the inflection points to form the boundary of the region with high probability/value.

---

Examples

Classification — Iris

import matplotlib.pyplot as plt
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression
from sheshe import ModalBoundaryClustering

iris = load_iris()
X, y = iris.data, iris.target

sh = ModalBoundaryClustering(
    base_estimator=LogisticRegression(max_iter=1000),
    task="classification",
    base_2d_rays=24,
    random_state=0,
    drop_fraction=0.5,
).fit(X, y)

print(sh.interpretability_summary(iris.feature_names).head())
sh.plot_pairs(X, y, max_pairs=3)   # generate the plots
plt.show()

Classification with pre-trained model

import matplotlib.pyplot as plt
from sklearn.datasets import load_wine
from sklearn.ensemble import RandomForestClassifier
from sheshe import ModalBoundaryClustering

wine = load_wine()
X, y = wine.data, wine.target

# Train a model independently
base_model = RandomForestClassifier(n_estimators=200, random_state=0)
base_model.fit(X, y)

# Use SheShe with that pre-fitted model
sh = ModalBoundaryClustering(
    base_estimator=base_model,
    task="classification",
    base_2d_rays=24,
    random_state=0,
    drop_fraction=0.5,
).fit(X, y)

sh.plot_pairs(X, y, max_pairs=2)
plt.show()

Classification — synthetic blobs with custom parameters

from sklearn.datasets import make_blobs
from sklearn.linear_model import LogisticRegression
from sheshe import ModalBoundaryClustering

X, y = make_blobs(n_samples=400, centers=5, cluster_std=1.8, random_state=0)

sh = ModalBoundaryClustering(
    base_estimator=LogisticRegression(max_iter=200),
    task="classification",
    base_2d_rays=16,
    scan_steps=32,
    n_max_seeds=3,
    direction="outside_in",
    random_state=0,
    drop_fraction=0.5,
).fit(X, y)

print(sh.predict(X[:5]))

Regression — Diabetes

import matplotlib.pyplot as plt
from sklearn.datasets import load_diabetes
from sklearn.ensemble import GradientBoostingRegressor
from sheshe import ModalBoundaryClustering

diab = load_diabetes()
X, y = diab.data, diab.target

sh = ModalBoundaryClustering(
    base_estimator=GradientBoostingRegressor(random_state=0),
    task="regression",
    base_2d_rays=24,
    random_state=0,
    drop_fraction=0.5,
).fit(X, y)

print(sh.interpretability_summary(diab.feature_names).head())
sh.plot_pairs(X, max_pairs=3)
plt.show()

Meta-optimization — Random search

The examples/meta_optimization.py script showcases a gradient-free meta- optimization routine. It evaluates random hyperparameter configurations for ModalBoundaryClustering directly instead of relying on first-order approximations, providing a simple way to tune the algorithm.

---

Benchmark

The percentile-based stopping rule avoids the point of inflection and scans only until the value crosses into a lower percentile bin (20 bins by default). The optimized loop implementation is considerably faster than the previous vectorized version. On the Iris dataset:

$ PYTHONPATH=src python experiments/benchmark_stop_criteria.py
vectorized implementation: 0.0259s
loop implementation:       0.0121s
speedup: 2.14x
ModalBoundaryClustering fit with stop_criteria='inflexion': 0.1026s
ModalBoundaryClustering fit with stop_criteria='percentile': 0.1411s

The exact numbers depend on the machine, but the optimized loop method is substantially quicker while producing the same results.

Saving figures

from pathlib import Path
import matplotlib.pyplot as plt

# after calling ``sh.plot_pairs(...)``
out_dir = Path("images")
out_dir.mkdir(exist_ok=True)
for i, fig_num in enumerate(plt.get_fignums()):
    fig = plt.figure(fig_num)
    fig.savefig(out_dir / f"pair_{i}.png")
    plt.close(fig)

Plotting with pandas DataFrames

import pandas as pd
import matplotlib.pyplot as plt
from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering

iris = load_iris()
df = pd.DataFrame(iris.data, columns=iris.feature_names)

sh = ModalBoundaryClustering().fit(df, iris.target)
sh.plot_pairs(df, iris.target, max_pairs=2)  # usa nombres de columnas en los ejes
plt.show()

Visualizing interpretability summary

import matplotlib.pyplot as plt

summary = sh.interpretability_summary(df.columns)
centroids = summary[summary["Type"] == "centroid"]
plt.scatter(centroids["coord_0"], centroids["coord_1"], c=centroids["Category"])
plt.xlabel("coord_0")
plt.ylabel("coord_1")
plt.show()

Save and load model

from pathlib import Path
from sklearn.datasets import load_iris
from sheshe import ModalBoundaryClustering

iris = load_iris()
X, y = iris.data, iris.target

sh = ModalBoundaryClustering().fit(X, y)
path = Path("sheshe_model.joblib")
sh.save(path)
sh2 = ModalBoundaryClustering.load(path)
print((sh.predict(X) == sh2.predict(X)).all())

For more complete examples, see the examples/ folder.

SubspaceScout

SubspaceScout helps discover informative feature subspaces (pairs, trios, ...) before running SheShe. It can work purely with mutual information or leverage optional models like LightGBM+SHAP or EBM to rank feature interactions.

from sheshe import SubspaceScout

scout = SubspaceScout(
    # model_method='lightgbm',    # default uses MI; LightGBM and SHAP are optional
    max_order=4,                # explore pairs, trios and quartets
    top_m=50,                   # limit to top 50 informative features
    base_pairs_limit=12,        # seed pairs for orders >=3
    beam_width=10,              # combos kept per layer
    extend_candidate_pool=16,   # random candidate features per parent
    branch_per_parent=4,        # extensions per parent
    marginal_gain_min=1e-3,     # minimum gain to accept
    max_eval_per_order=150,     # cap MI evaluations per order
    sample_size=4096,           # subsample size
    time_budget_s=None,         # e.g., 15.0 for 15 seconds
    task='classification',
    random_state=0,
)
subspaces = scout.fit(X, y)

ModalScoutEnsemble

ModalScoutEnsemble trains multiple ModalBoundaryClustering models on the top subspaces returned by SubspaceScout and combines their predictions.

from sheshe import ModalScoutEnsemble
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression

iris = load_iris()
X, y = iris.data, iris.target

mse = ModalScoutEnsemble(
    base_estimator=LogisticRegression(max_iter=200),
    task="classification",
    random_state=0,
    scout_kwargs={"max_order": 2, "top_m": 4, "sample_size": None},
    cv=2,
)
mse.fit(X, y)
print(mse.predict(X[:5]))

predict_proba(X)

Only available for classification tasks, this method returns the weighted mixture of class probabilities from all submodels in the ensemble.

mse.fit(X, y)
print(mse.predict_proba(X[:5]))

predict_regions(X)

Return the predicted label and cluster identifier for each sample.

labels, cluster_ids = mse.predict_regions(X[:3])
print(cluster_ids)

report()

report() returns a list with one entry per trained subspace, sorted by weight. Each entry is a dictionary containing:

• features: tuple with the indices of the features in that subspace.
• order: number of features (subspace order).
• scout_score: score assigned by SubspaceScout.
• cv_score: cross-validation score of the submodel.
• feat_importance: mean feature importance for the subspace.
• weight: normalized weight used by the ensemble.

Example:

from pprint import pprint

summary = mse.report()
pprint([
    {k: row[k] for k in ("features", "order", "scout_score", "cv_score", "feat_importance", "weight")}
    for row in summary[:2]
])

Output:

[{'cv_score': 0.9267,
  'feat_importance': 5.9886,
  'features': (3, 1),
  'order': 2,
  'scout_score': -0.2368,
  'weight': 0.4336},
 {'cv_score': 0.8467,
  'feat_importance': 7.3800,
  'features': (2, 1),
  'order': 2,
  'scout_score': -0.1543,
  'weight': 0.4193}]

plot_pairs(X, y=None, model_idx=0, max_pairs=None)

Visualize 2D decision surfaces of a given submodel using the same plotting utilities as ModalBoundaryClustering.

feats = mse.features_[0]
mse.plot_pairs(X, y, model_idx=0, max_pairs=1)

plot_pair_3d(X, pair, model_idx=0, class_label=None, grid_res=50, alpha_surface=0.6, engine="matplotlib")

Render probability (classification) or predicted value (regression) as a 3D surface for a specific submodel.

feats = mse.features_[0]
mse.plot_pair_3d(X, (feats[0], feats[1]), model_idx=0, class_label=mse.classes_[0])

Experiments and benchmark

The experiments comparing against unsupervised algorithms are located in the experiments/ folder. The script compare_unsupervised.py evaluates eight different datasets (Iris, Wine, Breast Cancer, Digits, California Housing, Moons, Blobs, Circles), explores parameters of SheShe, KMeans and DBSCAN, and stores four metrics (ARI, homogeneity, completeness, v_measure) along with the execution time (runtime_sec).

python experiments/compare_unsupervised.py --runs 5
cat benchmark/unsupervised_results_summary.csv | head

Results are generated inside benchmark/ (valores por repetición y medias en *_summary.csv).

An additional A/B comparison for the subspace-guided search is available in benchmark/subspace_ab_results.csv; the table below reports mean runtimes in seconds (5 seeds).

dataset	baseline	subspace	subspace + light + escape
digits	0.0567	0.0233	0.0222
iris	0.0040	0.00262	0.00268
=======
The new grad ray mode replaces the former grid approach, delivering up to
~9× faster fits with identical accuracy (see
benchmark/README.md).

For the manuscript we provide additional scripts in paper_experiments.py which perform supervised comparisons, ablation studies over base_2d_rays, direction, jaccard_threshold, drop_fraction and smooth_window, and sensitivity analyses w.r.t. dimensionality and Gaussian noise. Executing the script generates tables with todas las repeticiones y un resumen (*_summary.csv), además de figuras (*.png) bajo benchmark/:

python experiments/paper_experiments.py --runs 5

---

Key parameters

• base_2d_rays → controls angular resolution in 2D (32 by default). 3D scales to ~34; d>3 uses subspaces.
• direction → "center_out" | "outside_in" to locate the inflection point.
• scan_radius_factor, scan_steps → size and resolution of the radial scan.
• optim_method → "gradient_ascent" (default) or "trust_region_newton"; the trust-region variant uses gradients and Hessians to solve quadratic subproblems inside an adaptive radius and respects box constraints.
• grad_* → hyperparameters of gradient ascent (rate, iterations, tolerances; used only when optim_method="gradient_ascent").
• max_subspaces → max number of subspaces considered when d>3.
• density_alpha / density_k → optional density penalty computed with an HNSW k‑NN search (via hnswlib) to keep centers inside the data cloud. The normalized value is multiplied by (density(x))**density_alpha; set density_alpha=0 to disable.

---

Performance tips

• Defaults favour speed: base_2d_rays=32, scan_steps=24 and n_max_seeds=2.
• The heuristic auto_rays_by_dim=True (default) reduces rays for high dimensional datasets:
  • 25–64 features → base_2d_rays capped at 16.
  • 65+ features → base_2d_rays capped at 12. For 30D problems such as Breast Cancer this matches the recommended base_2d_rays=16.

---

Limitations

• Depends on the surface produced by the base model (can be rough in RF).
• In high dimension, the boundary is an approximation (subspaces).
• Finds local maxima (does not guarantee the global one), mitigated with multiple seeds.

---

Images

Figures have been intentionally omitted because this repository does not permit storing binary assets.

---

Contribute

Improvements are welcome. To propose changes:

1. Fork the repository and create a descriptive branch.

2. Install development dependencies and run the tests:

   pip install -e ".[dev]"
   PYTHONPATH=src pytest -q
   

3. Submit a pull request with a clear description of the change.

---

License

MIT