Publicação
BiAs Discovery and Mitigation in medical Artificial iNtelligence
| Resumo: | Machine-learning systems are used to improve efficiency and quality of results and should uphold an impartiality standard above human decisions. Nevertheless, biases are frequently observed, leading to suboptimal outcomes for specific groups. This problem is amplified in healthcare by the field’s complexity, limitations, and implications of its applications. The most common problem is the lack of enough samples to accurately represent the population, resulting in impeding consequences in these specific groups. Existing methods for evaluating these systems vary from evaluating the global per- formance of studied groups to comparing similar samples in an instance-based analysis. From the latter approach, the methodology of generating counterfactuals, samples modi- fied to answer "what if..?" scenarios, has gained popularity in recent years, valued for its interpretability and versatility. However, despite the instance-based perspective it provides, there is a gap in how to properly generalize this methodology. This work extends this approach by exploring novel evaluation metrics supported by a new visualization analogous to the confusion matrix. It also explores the plausibility of the generated counterfactuals, experimenting with the incorporation of domain knowledge. Motivated by a prevalent issue in healthcare - data scarcity - it analyzes the results of performing data augmentation with counterfactuals to mitigate bias without compromising performance. As a result, this work contributes with a new bias detection and mitigation technique and reports promising results for ensuring more reliable decision-support systems in healthcare. |
|---|---|
| Autores principais: | Pinto, Mariana Sofia Figueiredo |
| Assunto: | Bias Machine-Learning Counterfactuals Fairness Augmentation |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Machine-learning systems are used to improve efficiency and quality of results and should uphold an impartiality standard above human decisions. Nevertheless, biases are frequently observed, leading to suboptimal outcomes for specific groups. This problem is amplified in healthcare by the field’s complexity, limitations, and implications of its applications. The most common problem is the lack of enough samples to accurately represent the population, resulting in impeding consequences in these specific groups. Existing methods for evaluating these systems vary from evaluating the global per- formance of studied groups to comparing similar samples in an instance-based analysis. From the latter approach, the methodology of generating counterfactuals, samples modi- fied to answer "what if..?" scenarios, has gained popularity in recent years, valued for its interpretability and versatility. However, despite the instance-based perspective it provides, there is a gap in how to properly generalize this methodology. This work extends this approach by exploring novel evaluation metrics supported by a new visualization analogous to the confusion matrix. It also explores the plausibility of the generated counterfactuals, experimenting with the incorporation of domain knowledge. Motivated by a prevalent issue in healthcare - data scarcity - it analyzes the results of performing data augmentation with counterfactuals to mitigate bias without compromising performance. As a result, this work contributes with a new bias detection and mitigation technique and reports promising results for ensuring more reliable decision-support systems in healthcare. |
|---|