nettrustscore 2.0.2

Trustworthiness metrics for Softmax predictive models

Trustworthiness Quantification

This repository provides a Python implementation of trustworthiness quantification metrics for predictive models (e.g., DNNs):

The implementation is flexible and works out-the-box with any Python code that outputs SoftMax probabilities.

Key Features

• Question-Answer Trust: Computes trust scores per sample based on prediction correctness and softmax probability.
• Trust Density: Estimates the distribution of trust scores per class using Kernel Density Estimation (KDE).
• Trust Spectrum: Visualizes trust densities across all classes.
• NetTrustScore (NTS): Provides a scalar metric summarizing overall trustworthiness.

---

Installation

Recommended: Install via Conda-Forge

The easiest way to install nettrustscore is via Conda-Forge, which handles all dependencies automatically. Run the following command:

conda install -c conda-forge nettrustscore

Alternative: Manual Installation

If you prefer to install the package manually or are not using Conda, you can install the required dependencies and clone the repository.

Install Dependencies

• NumPy: For numerical calculations.
• Matplotlib: For plotting the trust spectrum.
• Scikit-learn: For Kernel Density Estimation (KDE) in trust density estimation.

Install them via pip:

pip install numpy matplotlib scikit-learn

or

Install them via conda:

conda install numpy matplotlib scikit-learn

Clone the Repository

git clone https://github.com/yaniker/nettrustscore-python.git
cd nettrustscore-python

Example Usage

from trustquant import NTS, CNTS #This is how the package is imported.
import numpy as np

# Example oracle and predictions
oracle = np.array([0, 0, 1, 2, 2, 0, 1])  # True labels
predictions = np.array([
    [0.8, 0.1, 0.1],  # Correct, high confidence
    [1.0, 0.0, 0.0],  # Correct, high confidence
    [0.2, 0.7, 0.1],  # Correct, high confidence
    [0.1, 0.2, 0.7],  # Correct, high confidence
    [0.1, 0.4, 0.5],  # Correct, lower confidence
    [0.1, 0.8, 0.1],  # Incorrect, high confidence
    [0.3, 0.3, 0.4]   # Incorrect, low confidence
]
) #Replace this with your model's predictions (`predictions = model.predict()`)

# FOR NETTRUSTSCORE #
# Initialize with default parameters
nts = NTS(oracle, predictions) #This is how you initialize. trust_spectrum = True will save trust spectrum to the directory under "trust_spectrum.png"
nts_scores_dict = nts.compute() # Computes trustworthiness for each class and overall.
print(nts_scores_dict)

# FOR CONDITIONAL NETTRUSTSCORE #
# Initialize with default parameters
cnts = CNTS(oracle, predictions) #This is how you initialize. trust_spectrum = True will save trust spectrum to the directory under "trust_spectrum.png" and "conditional_trust_densities.png"
cnts_scores_dict = cnts.compute() # Computes trustworthiness for each class and overall.
print(cnts_scores_dict)

Example Plot for Trust Spectrum (trust_spectrum = True)

Example Plot for Conditional Trust Spectrum (trust_spectrum = True)

I shared the codes for the plots Python scripts for plots for users to modify as needed.

This code was developed as a part of my publications listed below:

1. Yanik, E., Kruger, U., Intes, X., Rahul, R., & De, S. (2023). Video-based formative and summative assessment of surgical tasks using deep learning. Scientific Reports, 13(1), 1038.
2. Yanik, E., Ainam, J. P., Fu, Y., Schwaitzberg, S., Cavuoto, L., & De, S. (2024). Video-based skill acquisition assessment in laparoscopic surgery using deep learning. Global Surgical Education-Journal of the Association for Surgical Education, 3(1), 26.
3. Yanik, E., Schwaitzberg, S., Yang, G., Intes, X., Norfleet, J., Hackett, M., & De, S. (2024). One-shot skill assessment in high-stakes domains with limited data via meta learning. Computers in Biology and Medicine, 174, 108470. / Yanik, E., Schwaitzberg, S., Yang, J., Intes, X., & De, S. (2022). One-shot domain adaptation in video-based assessment of surgical skills. arXiv e-prints, arXiv-2301.
4. Yanik, E., Intes, X., & De, S. (2024). Cognitive-Motor Integration in Assessing Bimanual Motor Skills. arXiv preprint arXiv:2404.10889.

Licence

This project is licensed under the MIT License. See the file for details.

References for the methods:

1.How Much Can We Really Trust You? Towards Simple, Interpretable Trust Quantification Metrics for Deep Neural Networks.
2.Where Does Trust Break Down? A Quantitative Trust Analysis of Deep Neural Networks via Trust Matrix and Conditional Trust Densities.

If you use this package in your research, please cite the following papers accordingly:

For NTS:
@article{wong2020much,
  title={How much can we really trust you? towards simple, interpretable trust quantification metrics for deep neural networks},
  author={Wong, Alexander and Wang, Xiao Yu and Hryniowski, Andrew},
  journal={arXiv preprint arXiv:2009.05835},
  year={2020}
}

For conditional NTS:
@article{hryniowski2020does,
  title={Where does trust break down? A quantitative trust analysis of deep neural networks via trust matrix and conditional trust densities},
  author={Hryniowski, Andrew and Wang, Xiao Yu and Wong, Alexander},
  journal={arXiv preprint arXiv:2009.14701},
  year={2020}
}