casimac 1.2.2

Supervised multi-class/single-label classification with gradients

This Python project provides a supervised multi-class/single-label classification algorithm, which allows the prediction of class labels and their probabilities including gradients with respect to features. The classifier is designed along the principles of an scikit-learn estimator.

The details of the algorithm have been published in arXiv:2103.02926.

Complete documentation of the code is available via https://casimac.readthedocs.io/en/latest/.

Installation

Install the package via pip or clone this repository. In order to use pip, type:

$ pip install casimac

Getting Started

Use the CASIMAClassifier class to create a classifier object. This object provides a fit method for training and a predict method for the estimation of class labels. Furthermore, the predict_proba method can be used to predict class label probabilities.

Below is a short example.

from casimac import CASIMAClassifier

# Third party packages
import numpy as np
from sklearn.gaussian_process import GaussianProcessRegressor
import matplotlib.pyplot as plt

# Create data
N = 10
seed = 42
X = np.random.RandomState(seed).uniform(-10,10,N).reshape(-1,1)
y = np.zeros(X.size)
y[X[:,0]>0] = 1

# Classify
clf = CASIMAClassifier(GaussianProcessRegressor)
clf.fit(X, y)

# Predict
X_sample = np.linspace(-10,10,100).reshape(-1,1)
y_sample = clf.predict(X_sample)
p_sample = clf.predict_proba(X_sample)

# Plot result
plt.figure(figsize=(10,5))
plt.plot(X_sample,y_sample,label="class prediction")
plt.plot(X_sample,p_sample[:,1],label="class probability prediction")
plt.scatter(X,y,c='r',label="train data")
plt.xlabel("X")
plt.ylabel("label / probability")
plt.legend()
plt.show()

Also implemented are a fit_transform method and an inverse_transform method to map the latent variables to a unit simplex and vice versa. These methods work only on an already fitted classifier object. Example notebooks can be found in the examples directory.

📖 Citation

If you find this code useful, please consider citing:

@misc{heese2021calibrated,
      title={Calibrated Simplex Mapping Classification},
      author={Raoul Heese and Micha{\l} Walczak and Michael Bortz and Jochen Schmid},
      year={2021},
      eprint={2103.02926},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
     }

License

This project is licensed under the MIT License - see the LICENSE file for details.