Metadata-Version: 2.1
Name: scikit-psl
Version: 0.5.0
Summary: Probabilistic Scoring List classifier
Home-page: https://github.com/trr318/scikit-psl
License: MIT
Author: Stefan Heid
Author-email: stefan.heid@upb.de
Requires-Python: >=3.10,<3.13
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Requires-Dist: joblib (>=1.3.2,<2.0.0)
Requires-Dist: numpy (>=1.25.2,<2.0.0)
Requires-Dist: pandas (>=2.1.0,<3.0.0)
Requires-Dist: scikit-learn (>=1.3.0,<2.0.0)
Requires-Dist: scipy (>=1.11.2,<2.0.0)
Project-URL: Repository, https://github.com/trr318/scikit-psl
Description-Content-Type: text/markdown

[![License](https://img.shields.io/github/license/trr318/scikit-psl)](https://github.com/trr318/scikit-psl/blob/master/LICENSE)
[![Pip](https://img.shields.io/pypi/v/scikit-psl)](https://pypi.org/project/scikit-psl)
[![Paper](https://img.shields.io/badge/doi-10.1007%2F978--3--031--45275--8__13-green)](https://doi.org/10.1007/978-3-031-45275-8_13)


# Probabilistic Scoring Lists

Probabilistic scoring lists are incremental models that evaluate one feature of the dataset at a time.
PSLs can be seen as a extension to *scoring systems* in two ways:
- they can be evaluated at any stage allowing to trade of model complexity and prediction speed.
- they provide probablistic predictions instead of determnistic decisions for each possible score.

Scoring Systems are used as decision support for human experts in medical or law domains.

This implementation adheres to the [sklearn-api](https://scikit-learn.org/stable/glossary.html#glossary-estimator-types).

# Install
```bash
pip install scikit-psl
```

# Usage

For examples have a look at the `examples` folder, but here is a simple example


```python
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

from skpsl import ProbabilisticScoringList

# Generating synthetic data with continuous features and a binary target variable
X, y = make_classification(n_informative=10, random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.2, random_state=42)

psl = ProbabilisticScoringList({-1, 1, 2})
psl.fit(X_train, y_train)
print(f"Brier score: {psl.score(X_test, y_test):.4f}")
"""
Brier score: 0.2438  (lower is better)
"""

df = psl.inspect(5)
print(df.to_string(index=False, na_rep="-", justify="center", float_format=lambda x: f"{x:.2f}"))
"""
 Stage Threshold  Score  T = -2  T = -1  T = 0  T = 1  T = 2  T = 3  T = 4  T = 5
  0            -     -       -       -   0.51      -      -      -      -      - 
  1     >-2.4245  2.00       -       -   0.00      -   0.63      -      -      - 
  2     >-0.9625 -1.00       -    0.00   0.00   0.48   1.00      -      -      - 
  3      >0.4368 -1.00    0.00    0.00   0.12   0.79   1.00      -      -      - 
  4     >-0.9133  1.00    0.00    0.00   0.12   0.12   0.93   1.00      -      - 
  5      >2.4648  2.00    0.00    0.00   0.07   0.07   0.92   1.00   1.00   1.00 
"""
```

# Changelog

All notable changes to this project will be documented in this file.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

## 0.5.0 - 2023-11-16

### Added

- _ClassifierAtK
    - Sigmoid calibration additional to isotonic
- PSL classifier
    - Make optimization loss configurable
    - Small `searchspace_analyisis(·)` function makes lookahead choice more informed

### Fixed

- Fixed lookahead search space and considering global loss for model-sequence evaluation

### Changed

- Updated dependencies and added black
- Moved Binarizer to different module
- Moved PSL hyperparameters to constructor

## 0.4.2 - 2023-11-09

### Fixed

- _ClassifierAtK
    - Expected entropy for stage 0 now also calculated wrt. base 2
    - Data with only 0 or 1 is now also interpret as binary data

## 0.4.1 - 2023-10-19

### Fixed

- Small import error

## 0.4.0 - 2023-10-17

### Added

- Add brute force threshold optimization method to find the global optimum, bisect optimizer remains default method

### Changed

- Restructured source files

## 0.3.1 - 2023-09-12

### Fixed

- PSL is now correctly handles when all instances belong to the negative class
- [#1](../../issues/1) if the first feature is assigned a negative score, it is now assigned the most negative score

## 0.3.0 - 2023-08-10

### Added

- PSL classifier can now run with continuous data and optimally (wrt. expected entropy) select thresholds to binarize
  the data

### Changed

- Significantly improved optimum calculation for MinEntropyBinarizer (the same optimization algorithm is shared with the
  psls internal binarization algorithm)

## 0.2.0 - 2023-08-10

### Added

- PSL classifier
    - introduced parallelization
    - implemented l-step lookahead
    - simple inspect(·) method that creates a tabular representation of the model

## 0.1.0 - 2023-08-08

### Added

- Initial implementation of the PSL algorithm

