Tiago de Melo

Does SBBD mirror the global database research agenda or cultivate a distinct, locally shaped identity? We study this using large-scale topic modeling of a decade of publications: 489 SBBD and 3,036 VLDB full-text papers from 2016–2025. Applying BERTopic with multilingual embeddings and LLM-assisted label refinement, we extract 31 SBBD and 35 VLDB topics and align them semantically. We observe three patterns: (1) convergence, with 58% of SBBD topics matching VLDB topics, notably in data management, privacy, and NLP; (2) temporal lag, with SBBD adopting mainstream trends 1–4 years after their VLDB peak; and (3) divergence, with 13 SBBD-only topics highlighting a strongly applied, Brazil-centric identity. We conclude with a forward-looking research agenda for SBBD structured around five strategic directions.

Este trabalho apresenta o PlayReviewsPT, um dataset público com 97,2 milhões de avaliações em português brasileiro de 96 aplicativos do Google Play (2010–2025). Os dados são disponibilizados em formato JSONL, particionados por aplicativo, com metadados estruturados e o código do coletor. O artigo descreve a metodologia de coleta, o esquema dos dados e uma caracterização inicial. As análises mostram concentração em poucos aplicativos, forte polarização das notas, predominância de comentários curtos e crescimento expressivo do volume ao longo do tempo. O PlayReviewsPT é um recurso em larga escala para processamento de linguagem natural, mineração de avaliações de aplicativos e engenharia de software empírica.

Este trabalho aborda o problema de cold start em sistemas de recomendação, que ocorre quando há pouca ou nenhuma informação sobre novos produtos. Propõe-se a construção de Product Knowledge Graphs (PKGs) enriquecidos com opiniões extraídas das avaliações dos usuários utilizando Large Language Models (Sabiá-3.1), integrando as especificações dos produtos com informações subjetivas. O grafo enriquecido é processado para inferência de avaliações em cenários de item cold start: para cada categoria de aspecto, a polaridade inferida é a média aritmética das polaridades das tuplas de opinião de produtos similares que compartilham o mesmo atributo técnico no PKG. Os resultados mostram concordância de 88,9% na classificação de polaridade em relação à extração direta das avaliações do produto alvo.

Integrating large language models (LLMs) into data-intensive scientific workflows requires understanding their robustness, efficiency, and cross-provider consistency. This paper evaluates five LLMs (ChatGPT, Gemini, DeepSeek, Mistral, and Claude) as interchangeable components of a semantic enrichment pipeline, using 2,387 Billboard Year-End Hot 100 songs (2000–2023) as a case study. All models had near-perfect success rates (99.7–100%) but over 20-fold cost differences ($0.46–$10.70). Thematic agreement was substantial (Cohen’s Kappa = 0.70–0.81), and temporal trajectories were consistent across providers (Pearson r = 0.65–0.82). Thus, provider choice minimally affects aggregate analytical conclusions, allowing workflow designs that prioritize cost or latency without sacrificing reproducibility.

Online reviews play a crucial role in e-commerce, yet research on rating prediction for Brazilian Portuguese remains limited. This paper consolidates results from six interconnected studies investigating rating prediction and rating-text inconsistency detection. We evaluate approaches spanning classical machine learning with 58 textual features, BERT-based models, and ten large language models in zero-shot settings. Results show that BERTimbau achieves the best performance among fine-tuned models (MAE 0.56, RMSE 0.91), while DeepSeek and ChatGPT-4o lead among Large Language Models (LLMs) (RMSE 0.93). We also extend the analysis to a multilingual context with emoji signals across 13 European languages. For inconsistency detection, we find that LLM reliability varies substantially: ChatGPT-o3 shows low consistency across runs (κ = 0.18), while DeepSeek-3.2 achieves near-perfect agreement (κ > 0.95) with F1-score above 97%. Our findings provide practical guidelines for model selection based on accuracy requirements, training data availability, and cost constraints.

Ensuring soldering quality is critical in modern electronics manufacturing, particularly for advanced package-on-package (PoP) assemblies, where precise flux control directly influences joint reliability and component alignment. Conventional manual inspections, performed sporadically throughout production shifts, are insufficient to detect critical fluctuations in flux quantity, potentially leading to solderability defects and unscheduled downtimes. This study proposes a real-time monitoring system based on computer vision techniques using the OpenCV library to continuously evaluate the flux level at the dipping unit of pick-and-place machines. Images were preprocessed, and the flux fill rate within a Region of Interest was analyzed automatically. Despite the increasing use of automated visual systems in SMT lines, real-time monitoring of solder flux levels remains largely manual and inconsistent, leading to quality issues and rework. The proposed system was validated over a one-week industrial evaluation by performing 3578 inspections. The results demonstrated a classification accuracy of 98.21 percent, false failure rate of 1.79 percent, and Cohen's Kappa coefficient of 0.865, indicating excellent predictive agreement. By enabling the early detection of flux shortages and minimizing reliance on manual interventions, the system significantly enhances the reliability of the soldering process and the overall manufacturing quality of PoP assemblies.