psolr 0.1.1

Metaheuristic-Optimized Logistic Regression (PSO-LR / MOLR)

PSO-LR: Metaheuristic-Optimized Logistic Regression (MOLR)

---

Overview

PSO-LR is an open-source, sklearn-compatible implementation of Metaheuristic-Optimized Logistic Regression (MOLR).

Unlike traditional logistic regression, which relies on gradient-based solvers
(LBFGS, Newton-CG, SGD), PSO-LR estimates model parameters using
Particle Swarm Optimization (PSO) — a population-based global optimization algorithm.

This approach preserves the classical interpretability of logistic regression
while improving robustness, global convergence behavior, and stability in
noisy or ill-conditioned datasets.

---

Motivation

Traditional logistic regression can struggle when:

• Multicollinearity is present
• The optimization landscape is poorly conditioned
• Data contains strong noise
• Strong regularization is required

PSO-LR replaces gradient descent with a global, population-based search mechanism,
making it suitable for:

• Marketing analytics
• Choice modeling
• Behavioral modeling
• Econometrics
• Healthcare analytics
• High-noise real-world datasets

---

Key Features

• Binary Logistic Regression
• Multinomial Logistic Regression
• Particle Swarm Optimization (PSO)
• Early stopping for efficiency
• L1 / L2 regularization
• Hard coefficient constraints
• Fully sklearn-compatible API
• Interpretable coefficients & odds ratios
• Domain-independent design

---

Quick Start Example

from psolr import PSOLogisticRegression
from sklearn.datasets import make_classification

X, y = make_classification(n_samples=1000, n_features=10)

model = PSOLogisticRegression(
    pop_size=50,
    max_iter=200,
    random_state=42
)

model.fit(X, y)
y_pred = model.predict(X)
y_prob = model.predict_proba(X)

---

Mathematical Formulation

The logistic model remains unchanged:

[ P(y=1|x) = \sigma(x^\top \beta) ]

The regularized negative log-likelihood is optimized using PSO:

[ \min_{\beta} ; -\mathcal{L}(\beta) + \lambda \Omega(\beta) ]

PSO searches the parameter space globally without requiring gradients, ensuring robust convergence.

This removes reliance on gradients and enables global exploration.

---

Installation

pip install psolr

Upgrade:

pip install --upgrade psolr

---

Example: Hyperparameter Optimization

from sklearn.model_selection import RandomizedSearchCV

param_dist = {
    "C": [0.1, 1, 10],
    "pop_size": [40, 60, 80],
    "max_iter": [200, 300]
}

search = RandomizedSearchCV(
    PSOLogisticRegression(random_state=42),
    param_distributions=param_dist,
    scoring="roc_auc",
    cv=5
)

search.fit(X, y)

---

Use Cases

PSO-LR is particularly beneficial when:

• Interpretability is critical
• Data is noisy or unstable
• Traditional solvers fail to converge reliably
• Hard constraints on coefficients are required
• Global optimization is preferred

---

Citation

If you use PSO-LR in your research, please cite:

Dutta, K. (2026). PSO-LR: Particle Swarm Optimization based Logistic Regression (Version 0.1.1). Zenodo. https://doi.org/10.5281/zenodo.18639831

BibTeX:

@software{dutta2026psolr,
  author = {Dutta, Koustav},
  title = {PSO-LR: Particle Swarm Optimization based Logistic Regression},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18639831}
}

---

License

MIT License.

---

Author

Koustav Dutta
PhD Researcher — Machine Learning, Artificial Intelligence & Neuralcomputing

---

Contributing

Contributions and improvements are welcome.
Please open an issue or submit a pull request.