Guilherme Fonseca
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Guilherme Fonseca has a total of 67 co-authors. Here are the top co-authors by number of shared publications:
- Leonardo Rocha
Federal University of São João del-Rei, São João del-Rei, Brazil
11 shared publications - Marcos André Gonçalves
Federal University of Minas Gerais, Belo Horizonte, Brazil
7 shared publications - Washington Cunha
Universidade Estadual de Campinas
4 shared publications - Gabriel Prenassi
UFSJ
3 shared publications - Ana Cláudia Machado
Federal University of São João del-Rei, São João del-Rei, Brazil
2 shared publications - Celso França
Federal University of Minas Gerais, Belo Horizonte, Brazil
2 shared publications
Latest Publications
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 495-508
Avaliação Sistemática de Text-to-SQL em Português: Um Benchmark Unificado Multi-Domínio
A tarefa Text-to-SQL tem avançado significativamente com o uso de Grandes Modelos de Linguagem (LLMs), mas sua avaliação em português ainda é limitada. Este trabalho propõe um benchmark unificado que consolida quatro bases públicas de diferentes domínios e avalia seis LLMs distintos a partir das métricas EX e CM sob um protocolo experimental padronizado. Os resultados mostram que, embora os modelos superem 70% de EX no Spider, o melhor desempenho no DATASUS alcança apenas 27,5%. A discrepância entre CM e EX também evidencia limitações de generalização composicional. Esses resultados sugerem que avaliações isoladas podem superestimar a capacidade dos modelos e destacam a necessidade de benchmarks integrados e multi-domínio.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 495-508
Avaliação Sistemática de Text-to-SQL em Português: Um Benchmark Unificado Multi-Domínio
A tarefa Text-to-SQL tem avançado significativamente com o uso de Grandes Modelos de Linguagem (LLMs), mas sua avaliação em português ainda é limitada. Este trabalho propõe um benchmark unificado que consolida quatro bases públicas de diferentes domínios e avalia seis LLMs distintos a partir das métricas EX e CM sob um protocolo experimental padronizado. Os resultados mostram que, embora os modelos superem 70% de EX no Spider, o melhor desempenho no DATASUS alcança apenas 27,5%. A discrepância entre CM e EX também evidencia limitações de generalização composicional. Esses resultados sugerem que avaliações isoladas podem superestimar a capacidade dos modelos e destacam a necessidade de benchmarks integrados e multi-domínio.
- Open Access
- 10 July 2026
- Published by Sociedade Brasileira de Computação in Revista Eletrônica de Iniciação Científica em Computação
- Vol. 24, 432-437
Um Estudo Comparativo de Estratégias de Seleção de Exemplos para In-Context Learning Aplicado à Classificação Automática de Texto com LLMs
A Classificação Automática de Texto com LLMs pode ser realizada via zero-shot (ZS) (baixo custo e menor desempenho) ou por fine-tuning (FT) (maior custo e maior desempenho). Este trabalho investiga o in-context learning (ICL) como alternativa intermediária, analisando o trade-off entre efetividade e eficiência. Avaliamos estratégias de seleção de exemplos, comparando abordagem aleatória e métodos baseados em representações vetoriais (TF-IDF, RoBERTa, SBERT e LLM2Vec), além da variação do número de exemplos nos prompts. Os resultados indicam que boas estratégias de seleção tornam o ICL promissor, superando o ZS em efetividade com custo inferior ao FT. Ainda assim, o melhor custo-benefício permanece com o FT de SLMs como o RoBERTa.
- 10 July 2026
- Published by Sociedade Brasileira de Computação in Revista Eletrônica de Iniciação Científica em Computação
- Vol. 24, 432-437
Open AccessUm Estudo Comparativo de Estratégias de Seleção de Exemplos para In-Context Learning Aplicado à Classificação Automática de Texto com LLMs
A Classificação Automática de Texto com LLMs pode ser realizada via zero-shot (ZS) (baixo custo e menor desempenho) ou por fine-tuning (FT) (maior custo e maior desempenho). Este trabalho investiga o in-context learning (ICL) como alternativa intermediária, analisando o trade-off entre efetividade e eficiência. Avaliamos estratégias de seleção de exemplos, comparando abordagem aleatória e métodos baseados em representações vetoriais (TF-IDF, RoBERTa, SBERT e LLM2Vec), além da variação do número de exemplos nos prompts. Os resultados indicam que boas estratégias de seleção tornam o ICL promissor, superando o ZS em efetividade com custo inferior ao FT. Ainda assim, o melhor custo-benefício permanece com o FT de SLMs como o RoBERTa.
- Open Access
- 24 June 2026
- Published by JMIR Publications Inc. in JMIR Formative Research
- Vol. 10, e86379
Beyond Area Under the Receiver Operating Characteristic Curve: Evaluating Predictive Performance Metrics Under Class Imbalance in Real-World Clinical Data
Background: Predictive models increasingly support clinical decision-making, although imbalanced outcome distributions are common in health care datasets and can distort performance evaluation. The area under the receiver operating characteristic curve (AUROC) remains the most frequently reported metric, despite its limited ability to reflect clinically meaningful performance under class imbalance. Objective: This study aimed to examine the influences of metric selection on the clinical interpretation of predictive models in imbalanced real-world health care data. Methods: This was a retrospective cohort study, including 17,018 hospitalized patients with COVID-19. Two predictive models using extreme gradient boosting (XGBoost) were developed to predict kidney replacement therapy (KRT) and mortality. Model performance was assessed using AUROC, macro- F 1 -score, class-specific precision and recall, calibration (curve, slope, and intercept), decision curve analysis, and learning curves. Standard rebalancing strategies were applied exclusively to the training data to evaluate their impact on performance. Results: KRT occurred in 9.5%, and mortality in 18.0%. Although AUROC values were high (0.928 for KRT and 0.945 for mortality), performance in the minority class was substantially lower. For KRT, precision was 0.539 and recall 0.372; for mortality, precision was 0.725 and recall 0.718. Rebalancing strategies were associated with higher recall for the minority class, but this gain was accompanied by a reduction in precision, with minimal impact on AUROC values. As a result, AUROC remained high despite clinically relevant changes in error distribution between false positives and false negatives. The learning curves show a plateau-like shape, with stable validation performance across all training set sizes for both outcomes. Conclusions: AUROC alone is insufficient to evaluate prediction models in imbalanced health care scenarios, even with rebalancing. Routine reporting of class-aware metrics, alongside learning curve analysis, is essential to support robust and clinically meaningful evaluation of predictive models, rather than their direct translation into practice.
- 24 June 2026
- Published by JMIR Publications Inc. in JMIR Formative Research
- Vol. 10, e86379
Open AccessBeyond Area Under the Receiver Operating Characteristic Curve: Evaluating Predictive Performance Metrics Under Class Imbalance in Real-World Clinical Data
Background: Predictive models increasingly support clinical decision-making, although imbalanced outcome distributions are common in health care datasets and can distort performance evaluation. The area under the receiver operating characteristic curve (AUROC) remains the most frequently reported metric, despite its limited ability to reflect clinically meaningful performance under class imbalance. Objective: This study aimed to examine the influences of metric selection on the clinical interpretation of predictive models in imbalanced real-world health care data. Methods: This was a retrospective cohort study, including 17,018 hospitalized patients with COVID-19. Two predictive models using extreme gradient boosting (XGBoost) were developed to predict kidney replacement therapy (KRT) and mortality. Model performance was assessed using AUROC, macro- F 1 -score, class-specific precision and recall, calibration (curve, slope, and intercept), decision curve analysis, and learning curves. Standard rebalancing strategies were applied exclusively to the training data to evaluate their impact on performance. Results: KRT occurred in 9.5%, and mortality in 18.0%. Although AUROC values were high (0.928 for KRT and 0.945 for mortality), performance in the minority class was substantially lower. For KRT, precision was 0.539 and recall 0.372; for mortality, precision was 0.725 and recall 0.718. Rebalancing strategies were associated with higher recall for the minority class, but this gain was accompanied by a reduction in precision, with minimal impact on AUROC values. As a result, AUROC remained high despite clinically relevant changes in error distribution between false positives and false negatives. The learning curves show a plateau-like shape, with stable validation performance across all training set sizes for both outcomes. Conclusions: AUROC alone is insufficient to evaluate prediction models in imbalanced health care scenarios, even with rebalancing. Routine reporting of class-aware metrics, alongside learning curve analysis, is essential to support robust and clinically meaningful evaluation of predictive models, rather than their direct translation into practice.
- Open Access
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 103-111
Calibração de Sistemas de Recomendação com LLMs: Otimização de Prompts para Balancear Precisão, Diversidade e Justiça
Recommender Systems (RSs) play a central role in digital platforms, aiming to deliver relevant and personalized content. However, issues like popularity bias persist, limiting diversity and fairness. Calibration techniques seek to align recommendations with user preferences, often through post-processing adjustments. With the advent of Large Language Models (LLMs) such as GPT and LLaMA, new opportunities have emerged to personalize recommendations using prompt engineering. While recent approaches like prompt optimization have shown improvements in ranking accuracy, they often neglect other key aspects such as diversity, coverage, and fairness. This study investigates the use of LLMs for recommendation calibration, comparing their performance against traditional methods. We also evaluate the impact of different prompt optimization strategies across multiple metrics, including MAP, NDCG@10, MRMC, LTC, and GAP. Additionally, we employ a multicriteria utility function (MAUT) to analyze trade-offs between accuracy and diversity. Our results highlight the potential of LLMs and prompt engineering to enhance both personalization and fairness in Rss.
- 10 November 2025
- Published by Sociedade Brasileira de Computação
- p. 103-111
Open AccessCalibração de Sistemas de Recomendação com LLMs: Otimização de Prompts para Balancear Precisão, Diversidade e Justiça
Recommender Systems (RSs) play a central role in digital platforms, aiming to deliver relevant and personalized content. However, issues like popularity bias persist, limiting diversity and fairness. Calibration techniques seek to align recommendations with user preferences, often through post-processing adjustments. With the advent of Large Language Models (LLMs) such as GPT and LLaMA, new opportunities have emerged to personalize recommendations using prompt engineering. While recent approaches like prompt optimization have shown improvements in ranking accuracy, they often neglect other key aspects such as diversity, coverage, and fairness. This study investigates the use of LLMs for recommendation calibration, comparing their performance against traditional methods. We also evaluate the impact of different prompt optimization strategies across multiple metrics, including MAP, NDCG@10, MRMC, LTC, and GAP. Additionally, we employ a multicriteria utility function (MAUT) to analyze trade-offs between accuracy and diversity. Our results highlight the potential of LLMs and prompt engineering to enhance both personalization and fairness in Rss.
- Open Access
- 29 September 2025
- Published by Sociedade Brasileira de Computação
- p. 921-927
Um Estudo Comparativo de Estratégias de Seleção de Exemplos para In-Context Learning aplicado à Classificação Automática de Texto com Grandes Modelos de Linguagem
A Classificação Automática de Texto (CAT) com Grandes Modelos de Linguagem (LLMs) pode ser feita via zero-shot (baixo custo, menor efetividade) ou fine-tuning (alto custo, maior efetividade). Este estudo investiga o in-context learning, abordagem intermediária que insere poucos exemplos no prompt, avaliando estratégias de seleção de exemplos. Comparamos seleção aleatória com métodos baseados em representações vetoriais (TF-IDF, RoBERTa, SBERT e LLM2Vec). Os resultados reforçam que o fine-tuning de Small Language Models (SLMs), como RoBERTa, oferece melhor custo-benefício. Ainda assim, o incontext learning mostra-se promissor, superando o zero-shot em efetividade sem exigir os altos custos do fine-tuning, sobretudo com boas estratégias de seleção.
- 29 September 2025
- Published by Sociedade Brasileira de Computação
- p. 921-927
Open AccessUm Estudo Comparativo de Estratégias de Seleção de Exemplos para In-Context Learning aplicado à Classificação Automática de Texto com Grandes Modelos de Linguagem
A Classificação Automática de Texto (CAT) com Grandes Modelos de Linguagem (LLMs) pode ser feita via zero-shot (baixo custo, menor efetividade) ou fine-tuning (alto custo, maior efetividade). Este estudo investiga o in-context learning, abordagem intermediária que insere poucos exemplos no prompt, avaliando estratégias de seleção de exemplos. Comparamos seleção aleatória com métodos baseados em representações vetoriais (TF-IDF, RoBERTa, SBERT e LLM2Vec). Os resultados reforçam que o fine-tuning de Small Language Models (SLMs), como RoBERTa, oferece melhor custo-benefício. Ainda assim, o incontext learning mostra-se promissor, superando o zero-shot em efetividade sem exigir os altos custos do fine-tuning, sobretudo com boas estratégias de seleção.
- Open Access
- 19 May 2025
- Published by Association for Computing Machinery (ACM)
- p. 42-51
Characterizing YouTube's Role in Online Gambling Promotion: A Case Study of Fortune Tiger in Brazil
- 19 May 2025
- Published by Association for Computing Machinery (ACM)
- p. 42-51
Open AccessCharacterizing YouTube's Role in Online Gambling Promotion: A Case Study of Fortune Tiger in Brazil
- Open Access
- 27 September 2024
- Published by Springer Nature in BMC Infectious Diseases
- Vol. 24 (1) , 1-12
Comprehensive statistical analysis reveals significant benefits of COVID-19 vaccination in hospitalized patients: propensity score, covariate adjustment, and feature importance by permutation
COVID-19 vaccines effectively prevent infection and hospitalization. However, few population-based studies have compared the clinical characteristics and outcomes of patients hospitalized for COVID-19 using advanced statistical methods. Our objective is to address this evidence gap by comparing vaccinated and unvaccinated patients hospitalized for COVID-19. This retrospective cohort included adult COVID-19 patients admitted from March 2021 to August 2022 from 27 hospitals. Clinical characteristics, vaccination status, and outcomes were extracted from medical records. Vaccinated and unvaccinated patients were compared using propensity score analyses, calculated based on variables associated with vaccination status and/or outcomes, including waves. The vaccination effect was also assessed by covariate adjustment and feature importance by permutation. From the 3,188 patients, 1,963 (61.6%) were unvaccinated and 1,225 (38.4%) were fully vaccinated. Among these, 558 vaccinated individuals were matched with 558 unvaccinated ones. Vaccinated patients had lower rates of mortality (19.4% vs. 33.3%), invasive mechanical ventilation (IMV-18.3% vs. 34.6%), noninvasive mechanical ventilation (NIMV-10.6% vs. 22.0%), intensive care unit admission (ICU-32.0% vs. 44.1%) vasoactive drug use (21.1% vs. 32.6%), dialysis (8.2% vs. 14.7%) hospital length of stay (7.0 vs. 9.0 days), and thromboembolic events (3.9% vs.7.7%), p < 0.05 for all. Risk-adjusted multivariate analysis demonstrated a significant inverse association between vaccination and in-hospital mortality (adjusted odds ratio [aOR] = 0.42, 95% confidence interval [CI]: 0.31–0.56; p < 0.001) as well as IMV (aOR = 0.40, 95% CI: 0.30–0.53; p < 0.001). These results were consistent in all analyses, including feature importance by permutation. Vaccinated patients admitted to hospital with COVID-19 had significantly lower mortality and other severe outcomes than unvaccinated ones during the Delta and Omicron waves. These findings have important implications for public health strategies and support the critical importance of vaccination efforts, particularly in low-income countries, where vaccination coverage remains suboptimal.
- 27 September 2024
- Published by Springer Nature in BMC Infectious Diseases
- Vol. 24 (1) , 1-12
Open AccessComprehensive statistical analysis reveals significant benefits of COVID-19 vaccination in hospitalized patients: propensity score, covariate adjustment, and feature importance by permutation
COVID-19 vaccines effectively prevent infection and hospitalization. However, few population-based studies have compared the clinical characteristics and outcomes of patients hospitalized for COVID-19 using advanced statistical methods. Our objective is to address this evidence gap by comparing vaccinated and unvaccinated patients hospitalized for COVID-19. This retrospective cohort included adult COVID-19 patients admitted from March 2021 to August 2022 from 27 hospitals. Clinical characteristics, vaccination status, and outcomes were extracted from medical records. Vaccinated and unvaccinated patients were compared using propensity score analyses, calculated based on variables associated with vaccination status and/or outcomes, including waves. The vaccination effect was also assessed by covariate adjustment and feature importance by permutation. From the 3,188 patients, 1,963 (61.6%) were unvaccinated and 1,225 (38.4%) were fully vaccinated. Among these, 558 vaccinated individuals were matched with 558 unvaccinated ones. Vaccinated patients had lower rates of mortality (19.4% vs. 33.3%), invasive mechanical ventilation (IMV-18.3% vs. 34.6%), noninvasive mechanical ventilation (NIMV-10.6% vs. 22.0%), intensive care unit admission (ICU-32.0% vs. 44.1%) vasoactive drug use (21.1% vs. 32.6%), dialysis (8.2% vs. 14.7%) hospital length of stay (7.0 vs. 9.0 days), and thromboembolic events (3.9% vs.7.7%), p < 0.05 for all. Risk-adjusted multivariate analysis demonstrated a significant inverse association between vaccination and in-hospital mortality (adjusted odds ratio [aOR] = 0.42, 95% confidence interval [CI]: 0.31–0.56; p < 0.001) as well as IMV (aOR = 0.40, 95% CI: 0.30–0.53; p < 0.001). These results were consistent in all analyses, including feature importance by permutation. Vaccinated patients admitted to hospital with COVID-19 had significantly lower mortality and other severe outcomes than unvaccinated ones during the Delta and Omicron waves. These findings have important implications for public health strategies and support the critical importance of vaccination efforts, particularly in low-income countries, where vaccination coverage remains suboptimal.
- Open Access
- 1 January 2024
- Published by Sociedade Brasileira de Computação in Journal on Interactive Systems
- Vol. 15 (1) , 561-575
Pipelining Semantic Expansion and Noise Filtering for Sentiment Analysis of Short Documents – CluSent Method
The challenge of constructing effective sentiment models is exacerbated by a lack of sufficient information, particularly in short texts. Enhancing short texts with semantic relationships becomes crucial for capturing affective nuances and improving model efficacy, albeit with the potential drawback of introducing noise. This article introduces a novel approach, CluSent, designed for customized dataset-oriented sentiment analysis. CluSent capitalizes on the CluWords concept, a proposed powerful representation of semantically related words. To address the issues of information scarcity and noise, CluSent addresses these challenges: (i) leveraging the semantic neighborhood of pre-trained word embedding representations to enrich document representation and (ii) introducing dataset-specific filtering and weighting mechanisms to manage noise. These mechanisms utilize part-of-speech and polarity/intensity information from lexicons. In an extensive experimental evaluation spanning 19 datasets and five state-of-the-art baselines, including modern transformer architectures, CluSent emerged as the superior method in the majority of scenarios (28 out of 38 possibilities), demonstrating noteworthy performance gains of up to 14% over the strongest baselines.
- 1 January 2024
- Published by Sociedade Brasileira de Computação in Journal on Interactive Systems
- Vol. 15 (1) , 561-575
Open AccessPipelining Semantic Expansion and Noise Filtering for Sentiment Analysis of Short Documents – CluSent Method
The challenge of constructing effective sentiment models is exacerbated by a lack of sufficient information, particularly in short texts. Enhancing short texts with semantic relationships becomes crucial for capturing affective nuances and improving model efficacy, albeit with the potential drawback of introducing noise. This article introduces a novel approach, CluSent, designed for customized dataset-oriented sentiment analysis. CluSent capitalizes on the CluWords concept, a proposed powerful representation of semantically related words. To address the issues of information scarcity and noise, CluSent addresses these challenges: (i) leveraging the semantic neighborhood of pre-trained word embedding representations to enrich document representation and (ii) introducing dataset-specific filtering and weighting mechanisms to manage noise. These mechanisms utilize part-of-speech and polarity/intensity information from lexicons. In an extensive experimental evaluation spanning 19 datasets and five state-of-the-art baselines, including modern transformer architectures, CluSent emerged as the superior method in the majority of scenarios (28 out of 38 possibilities), demonstrating noteworthy performance gains of up to 14% over the strongest baselines.
- Open Access
- 23 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 198-207
A Survey on Review - Aware Recommendation Systems
Current advances in Deep Neural Networks have motivated recent studies to return their interest in Review-Aware Recommender Systems (RSs). In this sense, we employ a systematic mapping approach by selecting 56 papers published on the main vehicles of the area, such as RecSys, VLDB, SIGIR, and others. Reading these works and synthesising their achievements, we provide an updated picture of this field by highlighting relevant outcomes, contributions, and limitations. Especially, we have identified two huge limitations in the area. First, many of these works do not provide their code sources. Second, there is a prevalent tendency to prioritize accuracy above other quality dimensions, despite the consensus within the community to assess the practical effectiveness of RSs through many metrics. Both observations create constraints that restrict reproducibility and hinder straightforward comparisons in the field. Addressing this gap, this work also provides an open-source library that compiles the principal strategies proposed in the literature. Furthermore, we conducted a comprehensive evaluation through several datasets and metrics. Our findings indicate that while no single algorithm demonstrates absolute superiority, RSs based on neural networks have particularly exhibited the most competitive performance.
- 23 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 198-207
Open AccessA Survey on Review - Aware Recommendation Systems
Current advances in Deep Neural Networks have motivated recent studies to return their interest in Review-Aware Recommender Systems (RSs). In this sense, we employ a systematic mapping approach by selecting 56 papers published on the main vehicles of the area, such as RecSys, VLDB, SIGIR, and others. Reading these works and synthesising their achievements, we provide an updated picture of this field by highlighting relevant outcomes, contributions, and limitations. Especially, we have identified two huge limitations in the area. First, many of these works do not provide their code sources. Second, there is a prevalent tendency to prioritize accuracy above other quality dimensions, despite the consensus within the community to assess the practical effectiveness of RSs through many metrics. Both observations create constraints that restrict reproducibility and hinder straightforward comparisons in the field. Addressing this gap, this work also provides an open-source library that compiles the principal strategies proposed in the literature. Furthermore, we conducted a comprehensive evaluation through several datasets and metrics. Our findings indicate that while no single algorithm demonstrates absolute superiority, RSs based on neural networks have particularly exhibited the most competitive performance.