Mariana Cúri
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Mariana Cúri has a total of 34 co-authors. Here are the top co-authors by number of shared publications:
- Alneu de Andrade Lopes
Universidade de São Paulo
5 shared publications - Diego Minatel
University of São Paulo
5 shared publications - Antonio Rafael Sabino Parmezan
University of São Paulo
3 shared publications - Nícolas Roque dos Santos
UC Riverside
3 shared publications - Dylan Molenaar
University of Amsterdam
2 shared publications - Gabriel Couto Tabak
University of São Paulo, ICMC, São Carlos, Brazil
2 shared publications
Latest Publications
- Open Access
- 18 March 2026
- Published by Springer Nature in AI and Ethics
- Vol. 6 (2) , 206
FairDIF: debiasing classifiers with item response theory and differential item functioning
This article proposes FairDIF, a reference framework that leverages Item Response Theory (IRT) and Differential Item Functioning (DIF) concepts to promote group fairness in classification problems. IRT and DIF are distinct yet complementary mathematical tools widely used in educational assessments. IRT enables robust evaluations by modeling the probability of a correct response as a function of item properties and respondent ability, effectively capturing individual performance variability. In contrast, DIF identifies whether specific items systematically favor respondents from certain sociodemographic groups, regardless of their underlying ability levels, thereby revealing potential sources of bias. Our framework first assesses classifier predictions using an IRT model. Subsequently, FairDIF analyzes potential bias through the DIF area method. This process yields rich insights into classifier performance and fairness, as well as the relative difficulty of correctly classifying each example. We can apply FairDIF directly or extend it to create new fairness-aware methods. Here, we present multiple avenues for employing FairDIF in more responsible solutions and reformulate and evaluate three distinct and promising methods that use the proposed framework as their core component: (i) sample reweighting, (ii) model selection, and (iii) threshold tuning. Results demonstrate that these methods effectively balance predictive performance with key group-fairness criteria relative to their counterparts. These findings emphasize FairDIF’s potential to support fair model development across different stages of machine learning pipelines. Moreover, this work advances the integration of IRT and DIF into machine learning tasks by expanding their scope and providing a foundation for future research in this emerging area.
- 18 March 2026
- Published by Springer Nature in AI and Ethics
- Vol. 6 (2) , 206
Open AccessFairDIF: debiasing classifiers with item response theory and differential item functioning
This article proposes FairDIF, a reference framework that leverages Item Response Theory (IRT) and Differential Item Functioning (DIF) concepts to promote group fairness in classification problems. IRT and DIF are distinct yet complementary mathematical tools widely used in educational assessments. IRT enables robust evaluations by modeling the probability of a correct response as a function of item properties and respondent ability, effectively capturing individual performance variability. In contrast, DIF identifies whether specific items systematically favor respondents from certain sociodemographic groups, regardless of their underlying ability levels, thereby revealing potential sources of bias. Our framework first assesses classifier predictions using an IRT model. Subsequently, FairDIF analyzes potential bias through the DIF area method. This process yields rich insights into classifier performance and fairness, as well as the relative difficulty of correctly classifying each example. We can apply FairDIF directly or extend it to create new fairness-aware methods. Here, we present multiple avenues for employing FairDIF in more responsible solutions and reformulate and evaluate three distinct and promising methods that use the proposed framework as their core component: (i) sample reweighting, (ii) model selection, and (iii) threshold tuning. Results demonstrate that these methods effectively balance predictive performance with key group-fairness criteria relative to their counterparts. These findings emphasize FairDIF’s potential to support fair model development across different stages of machine learning pipelines. Moreover, this work advances the integration of IRT and DIF into machine learning tasks by expanding their scope and providing a foundation for future research in this emerging area.
- Open Access
- 31 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 890-898
A DIF-Driven Threshold Tuning Method for Improving Group Fairness
- 31 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 890-898
Open AccessA DIF-Driven Threshold Tuning Method for Improving Group Fairness
- 27 December 2024
- Published by Springer Nature in Behaviormetrika
- p. 1-24
An evolutionary neural architecture search for item response theory autoencoders
- 27 December 2024
- Published by Springer Nature in Behaviormetrika
- p. 1-24
An evolutionary neural architecture search for item response theory autoencoders
- 2 July 2024
- Published by Springer Nature
- p. 364-372
Masked Autoencoder Transformer for Missing Data Imputation of PISA
- 2 July 2024
- Published by Springer Nature
- p. 364-372
Masked Autoencoder Transformer for Missing Data Imputation of PISA
- 29 June 2024
- Published by Springer Nature
- p. 184-198
Item Response Theory in Sample Reweighting to Build Fairer Classifiers
- 29 June 2024
- Published by Springer Nature
- p. 184-198
Item Response Theory in Sample Reweighting to Build Fairer Classifiers
- Open Access
- 1 January 2024
- Published by EDITORA SCIENTIFIC in Brazilian Journal of Psychiatry
- Vol. 47, 1-27
Item-wise validity study of the Impact of Event Scale - Revised (IES-R): self-reported reactions of hospital personnel exposed to a stressful workplace
The Brazilian Journal of Psychiatry is the official publication of Brazilian Psychiatric Association (Associação Brasileira de Psiquiatria - ABP)
- 1 January 2024
- Published by EDITORA SCIENTIFIC in Brazilian Journal of Psychiatry
- Vol. 47, 1-27
Open AccessItem-wise validity study of the Impact of Event Scale - Revised (IES-R): self-reported reactions of hospital personnel exposed to a stressful workplace
The Brazilian Journal of Psychiatry is the official publication of Brazilian Psychiatric Association (Associação Brasileira de Psiquiatria - ABP)
- 27 November 2023
- Published by Springer Nature
- p. 630-645
DIF-SR: A Differential Item Functioning-Based Sample Reweighting Method
- 27 November 2023
- Published by Springer Nature
- p. 630-645
DIF-SR: A Differential Item Functioning-Based Sample Reweighting Method
- Open Access
- 23 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 144-153
Unfairness in Machine Learning for Web Systems Applications
- 23 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 144-153
Open AccessUnfairness in Machine Learning for Web Systems Applications
- Open Access
- 9 February 2023
- Published by Sociedade Brasileira de Computação in Revista Brasileira de Informática na Educação
- Vol. 31, 60-86
Teste Adaptativo Multiestágio para o ENEM
Testes internacionais de avaliação de alunos, em anos recentes, alteraram suas estruturas para implementar o formato adaptativo. O ENEM digital torna possível uma reestruturação da prova aderindo também aos testes adaptativos. Este artigo propõe um teste adaptativo multiestágio (TAM) para a prova do ENEM baseado na edição de 2019 na área de Matemática. Analisaram-se os itens do ENEM através da Teoria de Resposta ao Item, os quais foram utilizados para construir os módulos e estágios do TAM. O roteamento entre os módulos foi definido por um estudo do ponto de corte ótimo para o traço latente estimado, testado exaustivamente para encontrar o que trouxesse o melhor resultado comparado com a prova completa. Ao final, propôs-se uma arquitetura do teste multiestágio com valores de corte específicos para o roteamento. Constatou-se que o teste adaptativo reduziu em 44,4% o número de itens da prova e a estimação das habilidades foi mantida próxima da estimação com o teste completo. Notou-se também que o exame do Enem é voltado para a avaliação de níveis mais altos na escala de habilidade, tornando a estimação das habilidades prejudicada para indivíduos menos proficientes.
- 9 February 2023
- Published by Sociedade Brasileira de Computação in Revista Brasileira de Informática na Educação
- Vol. 31, 60-86
Open AccessTeste Adaptativo Multiestágio para o ENEM
Testes internacionais de avaliação de alunos, em anos recentes, alteraram suas estruturas para implementar o formato adaptativo. O ENEM digital torna possível uma reestruturação da prova aderindo também aos testes adaptativos. Este artigo propõe um teste adaptativo multiestágio (TAM) para a prova do ENEM baseado na edição de 2019 na área de Matemática. Analisaram-se os itens do ENEM através da Teoria de Resposta ao Item, os quais foram utilizados para construir os módulos e estágios do TAM. O roteamento entre os módulos foi definido por um estudo do ponto de corte ótimo para o traço latente estimado, testado exaustivamente para encontrar o que trouxesse o melhor resultado comparado com a prova completa. Ao final, propôs-se uma arquitetura do teste multiestágio com valores de corte específicos para o roteamento. Constatou-se que o teste adaptativo reduziu em 44,4% o número de itens da prova e a estimação das habilidades foi mantida próxima da estimação com o teste completo. Notou-se também que o exame do Enem é voltado para a avaliação de níveis mais altos na escala de habilidade, tornando a estimação das habilidades prejudicada para indivíduos menos proficientes.