Lucas G. L. Costa
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Lucas G. L. Costa has a total of 37 co-authors. Here are the top co-authors by number of shared publications:
- Gisele Lobo Pappa
Universidade Federal de Minas Gerais, Belo Horizonte, Brazil
14 shared publications - Gabriel Pereira de Oliveira
Universidade Federal de Minas Gerais (UFMG)
12 shared publications - Mariana O. Silva
UFMG
10 shared publications - Marco Túlio Dutra
UFMG / UFOP
8 shared publications - Anisio Lacerda
Universidade Federal de Minas Gerais, Belo Horizonte, Brazil
6 shared publications - Michele A. Brandão
UFMG
6 shared publications
Latest Publications
- Open Access
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 938-944
A Comparative Evaluation of IP Geolocation Services for Brazilian Data Enrichment
IP geolocation information is fundamental for enhancing analytical datasets, supporting applications such as recommendation systems, investigation, service personalization, among others digital fields. However, there are several geolocation discovery free online services as database enrichment tools. This work evaluate several of those services, aiming to define the most reliable service for the Brazilian network context, and compares their performance across IPv4 and IPv6. The ground truth set is extracted from Ripe Atlas’s brazilians devices. The experiment shows that BigData Cloud is the best service and have better results with IPv6 results. Despite the inherent limitation in obtaining reliable IP-based geolocation data, these findings help practitioners select more reliable services for geolocation enrichment of Brazilian IP datasets.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 938-944
Open AccessA Comparative Evaluation of IP Geolocation Services for Brazilian Data Enrichment
IP geolocation information is fundamental for enhancing analytical datasets, supporting applications such as recommendation systems, investigation, service personalization, among others digital fields. However, there are several geolocation discovery free online services as database enrichment tools. This work evaluate several of those services, aiming to define the most reliable service for the Brazilian network context, and compares their performance across IPv4 and IPv6. The ground truth set is extracted from Ripe Atlas’s brazilians devices. The experiment shows that BigData Cloud is the best service and have better results with IPv6 results. Despite the inherent limitation in obtaining reliable IP-based geolocation data, these findings help practitioners select more reliable services for geolocation enrichment of Brazilian IP datasets.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 603-616
Severe Name Homonymy: A Learning to Rank Approach for Civil Registry Entity Linking
This study addresses the challenge of transforming string-based kinship references into explicit entity relationships within large-scale civil registries. In scenarios where a record provides only the name of a relative (e.g., a mother’s name), identifying the correct unique identifier among numerous homonymous candidates constitutes a complex Known-Item Search problem. We propose a supervised Learning to Rank framework designed to disambiguate and rank these candidates by integrating kinship-based relational features, conditional probability signals, demographic indicators, and BERT-based semantic similarity. The approach was evaluated on a large-scale private dataset from Brazil and four public datasets: IPUMS (covering USA, Norway, and Iceland) and Wikidata. Experimental results demonstrate that our method provides a robust solution for identity resolution under high name ambiguity, achieving a Recall@5 above 0.90 across all evaluated datasets and enabling the construction of accurate entity links in civil databases.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 603-616
Severe Name Homonymy: A Learning to Rank Approach for Civil Registry Entity Linking
This study addresses the challenge of transforming string-based kinship references into explicit entity relationships within large-scale civil registries. In scenarios where a record provides only the name of a relative (e.g., a mother’s name), identifying the correct unique identifier among numerous homonymous candidates constitutes a complex Known-Item Search problem. We propose a supervised Learning to Rank framework designed to disambiguate and rank these candidates by integrating kinship-based relational features, conditional probability signals, demographic indicators, and BERT-based semantic similarity. The approach was evaluated on a large-scale private dataset from Brazil and four public datasets: IPUMS (covering USA, Norway, and Iceland) and Wikidata. Experimental results demonstrate that our method provides a robust solution for identity resolution under high name ambiguity, achieving a Recall@5 above 0.90 across all evaluated datasets and enabling the construction of accurate entity links in civil databases.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 677-682
Painel Trilhas de Licitações: Trilhas Investigativas para Geração de Alertas em Licitações de Minas Gerais
Este artigo apresenta o Painel Trilhas de Licitações, um sistema inteligente para apoio à detecção de indícios de fraude em licitações públicas no estado de Minas Gerais. A ferramenta integra múltiplas fontes de dados governamentais e utiliza trilhas investigativas para geração automática de alertas e priorização de auditorias. O sistema combina integração de dados com análise relacional para identificar conexões suspeitas entre licitantes e itens licitados, além de oferecer visualizações interativas que auxiliam auditores e promotores na investigação. Os resultados indicam potencial para aumentar a eficiência da fiscalização e fortalecer o combate à corrupção.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 677-682
Painel Trilhas de Licitações: Trilhas Investigativas para Geração de Alertas em Licitações de Minas Gerais
Este artigo apresenta o Painel Trilhas de Licitações, um sistema inteligente para apoio à detecção de indícios de fraude em licitações públicas no estado de Minas Gerais. A ferramenta integra múltiplas fontes de dados governamentais e utiliza trilhas investigativas para geração automática de alertas e priorização de auditorias. O sistema combina integração de dados com análise relacional para identificar conexões suspeitas entre licitantes e itens licitados, além de oferecer visualizações interativas que auxiliam auditores e promotores na investigação. Os resultados indicam potencial para aumentar a eficiência da fiscalização e fortalecer o combate à corrupção.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 89-96
Contradiction Detection in Brazilian Legislative Bills: A Limited Labeled Data Study
Identifying contradictory legislative bills is essential for ensuring consistency in Brazilian legal frameworks, yet it remains a manual and hard-to-scale task. In this work, we study automatic contradiction detection under limited labeled data. We use a dataset of 106 contradictory bill pairs curated by domain specialists and investigate two strategies: (i) supervised learning with synthetic data generation and (ii) zero-shot and few-shot classification using generative large language models (LLMs). Results show that few-shot LLMs perform strongly and outperform the supervised approach based on synthetic data. Although the small dataset limits generalization, our findings highlight the challenges of synthetic data in legal domains and the potential of LLMs for low-data scenarios.
- 8 September 2026
- Published by Sociedade Brasileira de Computação
- p. 89-96
Contradiction Detection in Brazilian Legislative Bills: A Limited Labeled Data Study
Identifying contradictory legislative bills is essential for ensuring consistency in Brazilian legal frameworks, yet it remains a manual and hard-to-scale task. In this work, we study automatic contradiction detection under limited labeled data. We use a dataset of 106 contradictory bill pairs curated by domain specialists and investigate two strategies: (i) supervised learning with synthetic data generation and (ii) zero-shot and few-shot classification using generative large language models (LLMs). Results show that few-shot LLMs perform strongly and outperform the supervised approach based on synthetic data. Although the small dataset limits generalization, our findings highlight the challenges of synthetic data in legal domains and the potential of LLMs for low-data scenarios.
- Open Access
- 5 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 189-199
ReSB²: Retrieving Similar Brazilian State Bills
- 5 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 189-199
Open AccessReSB²: Retrieving Similar Brazilian State Bills
- Open Access
- 1 January 2026
- Published by Sociedade Brasileira de Computação in Journal on Interactive Systems
- Vol. 17 (1) , 456-472
Detecting and Analyzing Anomalies in City Expenditure Time Series
This paper presents an enhanced approach for detecting and analyzing anomalies in public expenditure time series. The method combines statistical techniques and machine learning models to identify unusual spending behaviors and to rank expenditure items according to both the frequency of anomalies and the monetary impact involved. The statistically generated anomalies produce suspicion alerts of potential fraud, with the objective of prioritizing cases and directing human audit efforts more effectively. In this extended version, we introduce a finer-grained analysis that examines spending patterns at a more detailed level within the governmental budget structure, allowing the detection of irregularities that may not be visible under more aggregated views. The approach is validated on a real-world dataset comprising more than one million city expenditure records from the state of Minas Gerais, and the results demonstrate its ability to reveal irregularities that may remain hidden under higher levels of aggregation.
- 1 January 2026
- Published by Sociedade Brasileira de Computação in Journal on Interactive Systems
- Vol. 17 (1) , 456-472
Open AccessDetecting and Analyzing Anomalies in City Expenditure Time Series
This paper presents an enhanced approach for detecting and analyzing anomalies in public expenditure time series. The method combines statistical techniques and machine learning models to identify unusual spending behaviors and to rank expenditure items according to both the frequency of anomalies and the monetary impact involved. The statistically generated anomalies produce suspicion alerts of potential fraud, with the objective of prioritizing cases and directing human audit efforts more effectively. In this extended version, we introduce a finer-grained analysis that examines spending patterns at a more detailed level within the governmental budget structure, allowing the detection of irregularities that may not be visible under more aggregated views. The approach is validated on a real-world dataset comprising more than one million city expenditure records from the state of Minas Gerais, and the results demonstrate its ability to reveal irregularities that may remain hidden under higher levels of aggregation.
- Open Access
- 20 July 2025
- Published by Sociedade Brasileira de Computação
- p. 121-132
Detecção e Análise de Anomalias em Séries Temporais de Despesas Municipais
Este artigo apresenta uma abordagem para a detecção e análise de anomalias em séries temporais de despesas públicas. A abordagem emprega uma combinação de técnicas avançadas para detecção de anomalias em séries temporais, incluindo análise estatística e aprendizado de máquina. Além da detecção de anomalias, a abordagem permite o ranqueamento das despesas conforme a quantidade de anomalias identificadas e nos valores envolvidos, facilitando a priorização das auditorias. A abordagem é avaliada em um conjunto de dados reais contendo mais de um milhão de registros de despesas municipais do estado de Minas Gerais. Os experimentos indicam que a abordagem é capaz de identificar e priorizar casos com alto potencial de irregularidade.
- 20 July 2025
- Published by Sociedade Brasileira de Computação
- p. 121-132
Open AccessDetecção e Análise de Anomalias em Séries Temporais de Despesas Municipais
Este artigo apresenta uma abordagem para a detecção e análise de anomalias em séries temporais de despesas públicas. A abordagem emprega uma combinação de técnicas avançadas para detecção de anomalias em séries temporais, incluindo análise estatística e aprendizado de máquina. Além da detecção de anomalias, a abordagem permite o ranqueamento das despesas conforme a quantidade de anomalias identificadas e nos valores envolvidos, facilitando a priorização das auditorias. A abordagem é avaliada em um conjunto de dados reais contendo mais de um milhão de registros de despesas municipais do estado de Minas Gerais. Os experimentos indicam que a abordagem é capaz de identificar e priorizar casos com alto potencial de irregularidade.
- 30 January 2025
- Published by Springer Nature
- p. 19-32
GovBERT-BR: A BERT-Based Language Model for Brazilian Portuguese Governmental Data
- 30 January 2025
- Published by Springer Nature
- p. 19-32
GovBERT-BR: A BERT-Based Language Model for Brazilian Portuguese Governmental Data
- Open Access
- 1 January 2025
- Published by Sociedade Brasileira de Computação in Journal on Interactive Systems
- Vol. 16 (1) , 381-391
Fraud Alerts in Public Health Based on Audit Trails
Fraud detection and prevention in the public health sector are essential for preserving resource integrity and service quality. Fraud in this context can include financial misappropriations and data manipulation, which not only waste public funds but also compromise healthcare quality and accessibility. This article introduces an audit trail-based approach to identify and rank potential fraud cases involving public health employees in Brazil. Our ranking system highlights cases with higher fraud likelihood based on the analysis of suspicious activity patterns and travel records. Extending prior work, we also introduce three additional analyses on the alerts: temporal evolution, geographical distribution, and occupation-specific trends. Our results enable auditors to trace trends, regional irregularities, and occupation-specific risks, thus enhancing the accuracy and impact of fraud detection strategies.
- 1 January 2025
- Published by Sociedade Brasileira de Computação in Journal on Interactive Systems
- Vol. 16 (1) , 381-391
Open AccessFraud Alerts in Public Health Based on Audit Trails
Fraud detection and prevention in the public health sector are essential for preserving resource integrity and service quality. Fraud in this context can include financial misappropriations and data manipulation, which not only waste public funds but also compromise healthcare quality and accessibility. This article introduces an audit trail-based approach to identify and rank potential fraud cases involving public health employees in Brazil. Our ranking system highlights cases with higher fraud likelihood based on the analysis of suspicious activity patterns and travel records. Extending prior work, we also introduce three additional analyses on the alerts: temporal evolution, geographical distribution, and occupation-specific trends. Our results enable auditors to trace trends, regional irregularities, and occupation-specific risks, thus enhancing the accuracy and impact of fraud detection strategies.
- Open Access
- 14 October 2024
- Published by Sociedade Brasileira de Computação
- p. 247-259
Evaluating Domain-adapted Language Models for Governmental Text Classification Tasks in Portuguese
Domain-adaptive pre-training (DAPT) is a technique in natural language processing (NLP) that tailors pre-trained language models to specific domains, enhancing their performance in real-world applications. In this paper, we evaluate the effectiveness of DAPT in governmental text classification tasks, exploring how different factors, such as target domain dataset, pre-trained model language composition, and dataset size, impact model performance. We systematically vary these factors, creating distinct domain-adapted models derived from BERTimbau and LaBSE. Our experimental results reveal that selecting appropriate target domain datasets and pre-training strategies can notably enhance the performance of language models in governmental tasks.
- 14 October 2024
- Published by Sociedade Brasileira de Computação
- p. 247-259
Open AccessEvaluating Domain-adapted Language Models for Governmental Text Classification Tasks in Portuguese
Domain-adaptive pre-training (DAPT) is a technique in natural language processing (NLP) that tailors pre-trained language models to specific domains, enhancing their performance in real-world applications. In this paper, we evaluate the effectiveness of DAPT in governmental text classification tasks, exploring how different factors, such as target domain dataset, pre-trained model language composition, and dataset size, impact model performance. We systematically vary these factors, creating distinct domain-adapted models derived from BERTimbau and LaBSE. Our experimental results reveal that selecting appropriate target domain datasets and pre-training strategies can notably enhance the performance of language models in governmental tasks.