Periodic Reporting for period 2 - MAMMOth (Multi-Attribute, Multimodal Bias Mitigation in AI Systems)
Reporting period: 2024-05-01 to 2025-10-31
MAMMOth was a 36 – month Horizon Europe Research and Innovation Action, funded by the European Union under Grant Agreement ID: 101070285. MAMMOth developed an innovative fairness-aware AI framework that provides several tools and techniques for the discovery and mitigation of multi-discrimination, and for promoting the accountability of AI-systems with respect to multiple protected attributes and for tabular data and more complex network and visual data.
The MAMMOth consortium, which was coordinated by the Centre for Research and Technology Hellas (CERTH), includes computer scientists, AI experts, social scientists, public communication experts, ethics and data protection experts, as well as parties who represent communities of vulnerable and/or underrepresented groups in AI research.
MAMMOth made available both standalone open-source methods and an integrated open source “bias toolkit” (MAI-Bias) that combines new methods with third-party fairness libraries and components.
In addition to the developed research methods, algorithms and tools, the project engaged with communities of vulnerable and/or underrepresented groups in AI research (e.g. the LGBTIQ community, minority groups, migrants), implementing a co-creation strategy to ensure that genuine needs and pains are at the center of the research agenda. Furthermore, the multi-disciplinary approach of the project, which was supported by social science and ethics experts, grounded the project work in valid social science and humanity principles, and moved beyond a simplistic data-driven view of AI bias. It therefore contributed to the discovery of the possible underlying sources of bias and discrimination.
The MAMMOth tools were designed for three sectors of interest:
1. Algorithm-based decision making in finance: The goal was to identify attributes contributing to AI bias in credit scoring and debt repayment, and to develop and test an algorithmic decision-making system that reduces bias in financial services.
2. Decision making in face verification systems: The goal was to address inequalities in the access of minorities to online services using remote face verification, e.g. in the context of digital identity authentication/Know Your Customer (KYC) procedures.
3. Bias in academic collaborations and citations: The goal was to investigate how intersectional biases in search engines like Google Scholar affect the visibility of scholars and measure their impact on the academic network.
The MAMMOth consortium, which was coordinated by the Centre for Research and Technology Hellas (CERTH), includes computer scientists, AI experts, social scientists, public communication experts, ethics and data protection experts, as well as parties who represent communities of vulnerable and/or underrepresented groups in AI research.
MAMMOth made available both standalone open-source methods and an integrated open source “bias toolkit” (MAI-Bias) that combines new methods with third-party fairness libraries and components.
In addition to the developed research methods, algorithms and tools, the project engaged with communities of vulnerable and/or underrepresented groups in AI research (e.g. the LGBTIQ community, minority groups, migrants), implementing a co-creation strategy to ensure that genuine needs and pains are at the center of the research agenda. Furthermore, the multi-disciplinary approach of the project, which was supported by social science and ethics experts, grounded the project work in valid social science and humanity principles, and moved beyond a simplistic data-driven view of AI bias. It therefore contributed to the discovery of the possible underlying sources of bias and discrimination.
The MAMMOth tools were designed for three sectors of interest:
1. Algorithm-based decision making in finance: The goal was to identify attributes contributing to AI bias in credit scoring and debt repayment, and to develop and test an algorithmic decision-making system that reduces bias in financial services.
2. Decision making in face verification systems: The goal was to address inequalities in the access of minorities to online services using remote face verification, e.g. in the context of digital identity authentication/Know Your Customer (KYC) procedures.
3. Bias in academic collaborations and citations: The goal was to investigate how intersectional biases in search engines like Google Scholar affect the visibility of scholars and measure their impact on the academic network.
The project made significant progress across its scientific objectives, with a particular emphasis on the following:
1. Redefining Bias: The project advanced the understanding of bias by considering multiple protected characteristics, transcending the limitations of single-attribute fairness-aware learning. This multi-dimensional approach has been integrated into an operational framework that incorporates legal and societal perspectives to define and assess bias more comprehensively.
2. Standardised AI Solutions: A significant accomplishment of the project is the creation of the MAI-Bias toolkit and a set of libraries, all of which are open-source and provide standard APIs for measuring and mitigating group fairness or bias in multimodal AI systems. The toolkit and associated libraries provide an array of fairness building blocks that can be used to systematically explore bias and fairness, aiding in the development of more equitable AI systems.
3. Technology for Bias Evaluation and Mitigation: The project developed innovative methods (e.g. FairBranch, FLAC, BAdd, MAVias) to evaluate and mitigate bias, including quantifying bias under fuzzy logic. The research focuses on the belief values of discrimination and the possibility of encountering protected group members, offering a more nuanced and comprehensive assessment of AI fairness.
4. Reliability, Traceability, and Explainability: In line with the goal of ensuring reliable, traceable, and explainable AI solutions, the project leveraged Explainable Artificial Intelligence (XAI) methodologies and generative AI for creating equitable synthetic data, as well as studied the robustness of solutions under adversarial settings.
5. Availability, deployment and awareness raising of unbiased and bias-preventing AI solutions: By including experts and stakeholders in the MAMMOth bias toolkit's design, a co-creation process contributed to addressing the inherent application limits of technical and research work on AI fairness and encouraged awareness about MAMMOth topics among the research community and affected communities.
1. Redefining Bias: The project advanced the understanding of bias by considering multiple protected characteristics, transcending the limitations of single-attribute fairness-aware learning. This multi-dimensional approach has been integrated into an operational framework that incorporates legal and societal perspectives to define and assess bias more comprehensively.
2. Standardised AI Solutions: A significant accomplishment of the project is the creation of the MAI-Bias toolkit and a set of libraries, all of which are open-source and provide standard APIs for measuring and mitigating group fairness or bias in multimodal AI systems. The toolkit and associated libraries provide an array of fairness building blocks that can be used to systematically explore bias and fairness, aiding in the development of more equitable AI systems.
3. Technology for Bias Evaluation and Mitigation: The project developed innovative methods (e.g. FairBranch, FLAC, BAdd, MAVias) to evaluate and mitigate bias, including quantifying bias under fuzzy logic. The research focuses on the belief values of discrimination and the possibility of encountering protected group members, offering a more nuanced and comprehensive assessment of AI fairness.
4. Reliability, Traceability, and Explainability: In line with the goal of ensuring reliable, traceable, and explainable AI solutions, the project leveraged Explainable Artificial Intelligence (XAI) methodologies and generative AI for creating equitable synthetic data, as well as studied the robustness of solutions under adversarial settings.
5. Availability, deployment and awareness raising of unbiased and bias-preventing AI solutions: By including experts and stakeholders in the MAMMOth bias toolkit's design, a co-creation process contributed to addressing the inherent application limits of technical and research work on AI fairness and encouraged awareness about MAMMOth topics among the research community and affected communities.
The main project results include:
• A thorough study of AI bias with emphasis on multimodal and multi-attribute fairness, accompanied by numerous scientific publications in high impact venues.
• An open-source MAMMOth bias toolkit (MAI-Bias), which offers a user-friendly and versatile way for assessing and mitigating bias in complex datasets, with newly implemented algorithms, and easy-to-use library design with connections to popular bias libraries (e.g. AIF360).
• Establishment of best practices for incorporating bias-aware AI in credit scoring and face verification applications.
• Identification of the sources of biases in academic collaborations and citations (e.g. Google Scholar) and development of mitigation strategies for ranking algorithms.
• Training material about MAMMOth topics.
• A thorough study of AI bias with emphasis on multimodal and multi-attribute fairness, accompanied by numerous scientific publications in high impact venues.
• An open-source MAMMOth bias toolkit (MAI-Bias), which offers a user-friendly and versatile way for assessing and mitigating bias in complex datasets, with newly implemented algorithms, and easy-to-use library design with connections to popular bias libraries (e.g. AIF360).
• Establishment of best practices for incorporating bias-aware AI in credit scoring and face verification applications.
• Identification of the sources of biases in academic collaborations and citations (e.g. Google Scholar) and development of mitigation strategies for ranking algorithms.
• Training material about MAMMOth topics.