diff --git a/README.md b/README.md index e0c119b..d8018fa 100644 --- a/README.md +++ b/README.md @@ -3,9 +3,9 @@ 📄 Applied in real-world audit: [audit report](https://algorithmaudit.eu/algoprudence/cases/aa202402_preventing-prejudice_addendum/) -🌐 Web app: [Algorithm Audit website](https://algorithmaudit.eu/technical-tools/bdt/#web-app) +☁️ Web app on [Algorithm Audit website](https://algorithmaudit.eu/technical-tools/bdt/#web-app) -Note: The module is still considered experimental, so conclusions drawn from the results should be carefully reviewed by domain experts. +Note: This module is still considered experimental, so conclusions drawn from the results should be carefully reviewed by domain experts. ## Key takeaways – Why unsupervised bias detection? - **Quantitative-qualitative joint method**: Data-driven bias testing combined with the balanced and context-sensitive judgment of human experts; @@ -16,18 +16,21 @@ Note: The module is still considered experimental, so conclusions drawn from the - **Model-agnostic**: Works for all binary AI classifiers; - **Open-source and not-for-profit**: Easy to use and available for the entire AI auditing community. -|Overview|| -|---------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| -| **Code** | [![!pypi](https://img.shields.io/pypi/v/unsupervised-bias-detection?logo=pypi&color=blue)](https://pypi.org/project/unsupervised-bias-detection/) [![!python-versions](https://img.shields.io/pypi/pyversions/aeon?logo=python)](https://www.python.org/) [![license](https://img.shields.io/badge/license-MIT-blue)](https://github.com/NGO-Algorithm-Audit/unsupervised-bias-detection?tab=MIT-1-ov-file#)| -| **Community** | [![!slack](https://img.shields.io/static/v1?logo=slack&label=Slack&message=chat&color=lightgreen)](https://join.slack.com/t/aa-experthub/shared_invite/zt-2n8aqry8z-lWC6XTbqVmb6S2hpkThaqQ) [![!linkedin](https://img.shields.io/static/v1?logo=linkedin&label=LinkedIn&message=news&color=lightblue)](https://www.linkedin.com/company/algorithm-audit/)| -## Description of our Joint Fairness Assessment Method (JFAM) -The Joint Fairness Assessment Method developed by NGO Algorithm Audit combines data-driven bias testing with normative and context-sensitive judgment of human experts, to determine fair AI on a case-by-case basis. The data-driven component comprises an unsupervised clustering tool (available as a free-to-use [web application](https://algorithmaudit.eu/technical-tools/bdt/#web-app)) that discovers complex and hidden forms of bias. It thereby tackles the difficult problem of detecting proxy-discrimination that stems from unforeseen and higher-dimensional forms of bias, including intersectional forms of discrimination. The results of the bias scan tool serve as a starting point for a deliberative assessment by human experts to evaluate potential discrimination and unfairness in an AI system. -As an example, we applied our bias scan tool to a [BERT-based disinformation classifier](https://github.com/NGO-Algorithm-Audit/unsupervised-bias-detection/blob/master/classifiers/BERT_disinformation_classifier/BERT_Twitter_classifier.ipynb) and distilled a set of pressing questions about its performance and possible biases. We presented these questions to an independent advice commission composed of four academic experts on fair AI, and two civil society organizations working on disinformation detection. The advice commission believes there is a low risk of (higher-dimensional) proxy discrimination by the reviewed disinformation classifier. The commission judged that the differences in treatment identified by the quantitative bias scan can be justified, if certain conditions apply. The full advice can be read in our [algoprudence case repository](https://algorithmaudit.eu/algoprudence/cases/aa202301_bert-based-disinformation-classifier/). +| | | +| --- | --- | +| **Code** | [![!pypi](https://img.shields.io/pypi/v/unsupervised-bias-detection?logo=pypi&color=blue)](https://pypi.org/project/unsupervised-bias-detection/) [![!python-versions](https://img.shields.io/pypi/pyversions/aeon?logo=python)](https://www.python.org/) [![license](https://img.shields.io/badge/license-MIT-blue)](https://github.com/NGO-Algorithm-Audit/unsupervised-bias-detection?tab=MIT-1-ov-file#) | +| **Community** | [![!slack](https://img.shields.io/static/v1?logo=slack&label=Slack&message=chat&color=lightgreen)](https://join.slack.com/t/aa-experthub/shared_invite/zt-2n8aqry8z-lWC6XTbqVmb6S2hpkThaqQ) [![!linkedin](https://img.shields.io/static/v1?logo=linkedin&label=LinkedIn&message=news&color=lightblue)](https://www.linkedin.com/company/algorithm-audit/) | -Our joint approach to fair AI is supported by 20+ actors from the AI auditing community, including journalists, civil society organizations, NGOs, corporate data scientists and academics. In sum, it combines the power of rigorous, machine learning-informed bias testing with the balanced judgment of human experts, to determine fair AI in a concrete way. + +## How this tool fits in our quantitative-qualitative AI auditing framework? +The Joint Fairness Assessment Method developed (JFAM) by NGO Algorithm Audit combines data-driven bias testing with normative and context-sensitive judgment of human experts, to determine fair AI on a case-by-case basis. The data-driven component comprises this unsupervised clustering tool (available as a free-to-use [web app](https://algorithmaudit.eu/technical-tools/bdt/#web-app)) that discovers complex and hidden forms of bias. It thereby tackles the difficult problem of detecting proxy-discrimination that stems from unforeseen and higher-dimensional forms of bias, including intersectional forms of discrimination. The results of the bias scan tool serve as a starting point for a deliberative assessment by human experts to evaluate potential discrimination and unfairness in an AI system. + +As an example, we applied our bias detection tool to a [BERT-based disinformation classifier](https://github.com/NGO-Algorithm-Audit/unsupervised-bias-detection/blob/master/classifiers/BERT_disinformation_classifier/BERT_Twitter_classifier.ipynb) and distilled a set of pressing questions about its performance and possible biases. We presented these questions to an independent advice commission composed of four academic experts on fair AI, and two civil society organizations working on disinformation detection. The advice commission believes there is a low risk of (higher-dimensional) proxy discrimination by the reviewed disinformation classifier. The commission judged that the differences in treatment identified by the quantitative bias scan can be justified, if certain conditions apply. The full advice can be read in our [algoprudence case repository](https://algorithmaudit.eu/algoprudence/cases/aa202301_bert-based-disinformation-classifier/) (ALGO:AA:2023:01). + +Our joint approach to AI auditing is supported by 20+ actors from the international AI auditing community, including journalists, civil society organizations, NGOs, corporate data scientists and academics. In sum, it combines the power of rigorous, machine learning-informed bias testing with the balanced judgment of human experts, to determine fair AI in a concrete way. 1The bias scan tool is based on the k-means Hierarchical Bias-Aware Clustering method as described in Bias-Aware Hierarchical Clustering for detecting the discriminated groups of users in recommendation systems, Misztal-Radecka, Indurkya, _Information Processing and Management_ (2021). [[link]](https://www.sciencedirect.com/science/article/abs/pii/S0306457321000285) Additional research indicates that k-means HBAC, in comparison to other clustering algorithms, works best to detect bias in real-world datasets. @@ -40,24 +43,24 @@ Our joint approach to fair AI is supported by 20+ actors from the AI auditing co A .csv file of max. 1GB, with columns: features, predicted labels (named: 'pred_label'), ground truth labels (named: 'truth_label'). Note: Only the naming, not the order of the columns is of importance. The dataframe displayed in Table 1 is digestible by the web application. -| feat_1 | feat_2 | ... | feat_n | pred_label | truth_label | -|--------|--------|-----|--------|------------|-------------| -| 10 | 1 | ... | 0.1 | 1 | 1 | -| 20 | 2 | ... | 0.2 | 1 | 0 | -| 30 | 3 | ... | 0.3 | 0 | 0 | +| feat_1 | feat_2 | ... | feat_n | performance metric | +|--------|--------|-----|--------|------------| +| 10 | 1 | ... | 0.1 | 1 | +| 20 | 2 | ... | 0.2 | 1 | +| 30 | 3 | ... | 0.3 | 0 | *Table 1 – Structure of input data in the bias scan web application* -Note that the features values are unscaled numeric values, and 0 or 1 for the predicted and ground truth labels. +Features values can be numeric or categorical values. The numeric performance metric is context-dependent. The variable can, for instance, represents being 'selected for examination' (yes or no), 'assigned to a high-risk catagory (yes or no)' or false positive (yes or no). Low scores are considered to be a negative bias, i.e., if being selected for examination is considered to be harmful, 'selected for examination=Yes' should be codified as 0 and 'selected for examination=No' should be codified as 1. ## Contributing Members -- Floris | https://github.com/fholstege -- Joel Persson | https://github.com/jopersson -- Jurriaan | https://github.com/jfparie -- Kirtan | https://github.com/kirtanp -- Krsto | https://github.com/krstopro -- Mackenzie Jorgensen | https://github.com/mjorgen1 +- [Floris Holstege](https://github.com/fholstege) +- [Joel Persson](https://github.com/jopersson) +- [Jurriaan Parie](https://github.com/jfparie) +- [Kirtan Padh](https://github.com/kirtanp) +- [Krsto Proroković](https://github.com/krstopro) +- [Mackenzie Jorgensen](https://github.com/mjorgen1) ### 20+ endorsements from various parts of the AI auditing community #### Journalism @@ -86,6 +89,4 @@ Note that the features values are unscaled numeric values, and 0 or 1 for the pr - Marlies van Eck, Assistant Professor in Administrative Law & AI at Radboud University - Aileen Nielsen, Fellow Law&Tech at ETH Zürich - Vahid Niamadpour, PhD-candidate in Linguistics at Leiden University -- Ola Al Khatib, PhD-candidate in the legal regulation of algorithmic decision-making at Utrecht University -- Floris Holstege, PhD-candidate in Explainable Machine Learning at University of Amsterdam - +- Ola Al Khatib, PhD-candidate in the legal regulation of algorithmic decision-making at Utrecht University \ No newline at end of file