statsnb 0.2.0

Flexible Naive Bayes classifiers using scipy.stats distributions

StatsNB - Flexible Naive Bayes Classifiers

Naive Bayes classifiers using any scipy.stats continuous distribution, fully compatible with scikit-learn.

What is StatsNB?

GaussianNB from scikit-learn assumes Gaussian (normal) distributions. But what if your data follows a different distribution?

StatsNB lets you use all continuous distributions from scipy.stats, see scipy.stats documentation.

When to Use StatsNB?

Use StatsNB when:

• Your data is heavy-tailed (try t or cauchy distributions)
• Your data is bounded (try beta distribution)
• Your data is strictly positive and skewed (try gamma or lognorm)
• You want to experiment with different distributions

Installation

pip install statsnb

Quick Start

from statsnb import StatsNB
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

# Load data
X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

# Use t-distribution (more robust than Gaussian)
clf = StatsNB(dist='t')
clf.fit(X_train, y_train)
print(f"Accuracy: {clf.score(X_test, y_test):.3f}")

# Try different distributions
for dist in ['norm', 't', 'laplace', 'cauchy']:
    clf = StatsNB(dist=dist)
    clf.fit(X_train, y_train)
    print(f"{dist:10s}: {clf.score(X_test, y_test):.3f}")

Advanced Features

Fix Parameters During Fitting

# Fix degrees of freedom for t-distribution
clf = StatsNB(dist='t', fit_args={'fdf': 5})
clf.fit(X_train, y_train)

# Fix location for exponential distribution
clf = StatsNB(dist='expon', fit_args={'floc': 0})
clf.fit(X_train, y_train)

Custom Priors

# Specify class prior probabilities
clf = StatsNB(dist='gamma', priors=[0.2, 0.3, 0.5])
clf.fit(X_train, y_train)

Use Method of Moments (Faster)

# MM is much faster than MLE for some distributions
clf = StatsNB(dist='norm', fit_args={'method': 'MM'})
clf.fit(X_train, y_train)

API Compatibility

StatsNB is fully compatible with scikit-learn:

from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import GridSearchCV

# Use in pipelines
pipe = Pipeline([
    ('scaler', StandardScaler()),
    ('clf', StatsNB())
])
pipe.fit(X_train, y_train)

# Use in grid search
param_grid = {'clf__dist': ['norm', 't', 'laplace']}
grid = GridSearchCV(pipe, param_grid, cv=5)
grid.fit(X_train, y_train)
print(f"Best distribution: {grid.best_params_}")

Performance Note

Unlike GaussianNB which uses analytical solutions, StatsNB uses numerical optimization (MLE) to fit distributions. This makes it:

• ~5-10x slower than GaussianNB

• No incremental learning (partial_fit not supported)

  • Numerical optimization doesn't support online updates

• No sample weights

  • scipy.stats.fit() doesn't support weighted MLE

• More flexible - supports any distribution

• Still efficient for most real-world datasets

However, for norm,expon,gamma,laplace, parameter inference can be done with Methods of Momentum, thus support all features of GaussianNB.

Therefore, We provide MomentNB, which supports all GaussianMB features

• Incremental learning
• Sample weights
• Variance Smoothing
• And much faster!

However,MomentNB only supports norm,expon,gamma,laplace, setting dist= other distributions will result in an error.

Examples

See the examples/ directory for more.

Requirements

• Python ≥ 3.8
• numpy ≥ 1.20.0
• scipy ≥ 1.7.0
• scikit-learn ≥ 1.0.0

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

MIT License - see LICENSE file for details.

Citation

If you use StatsNB in your research, please cite:

@software{statsnb2025,
  author = {Your Name},
  title = {StatsNB: Flexible Naive Bayes Classifiers},
  year = {2025},
  url = {https://github.com/yourusername/statsnb}
}

Acknowledgments

• Built on top of scikit-learn and scipy
• Inspired by the flexibility needs of real-world data science