Álvaro Figueira

Abstract
This article focuses on the creation and evaluation of synthetic data to address the challenges of imbalanced datasets in machine learning (ML) applications, using fake news detection as a case study. We conducted a thorough literature review on generative adversarial networks (GANs) for tabular data, synthetic data generation methods, and synthetic data quality assessment. By augmenting a public news dataset with synthetic data generated by different GAN architectures, we demonstrate the potential of synthetic data to improve ML models' performance in fake news detection. Our results show a significant improvement in classification performance, especially in the underrepresented class. We also modify and extend a data usage approach to evaluate the quality of synthetic data and investigate the relationship between synthetic data quality and data augmentation performance in classification tasks. We found a positive correlation between synthetic data quality and performance in the underrepresented class, highlighting the importance of high-quality synthetic data for effective data augmentation.

Abstract
In today’s competitive higher education sector, institutions increasingly rely on international rankings to secure financial resources, attract top-tier talent, and elevate their global reputation. Simultaneously, these universities have expanded their presence on social media, utilizing sophisticated posting strategies to disseminate information and boost recognition and engagement. This study examines the relationship between higher education institutions’ (HEIs’) rankings and their social media posting strategies. We gathered and analyzed publications from 18 HEIs featured in a consolidated ranking system, examining various features of their social media posts. To better understand these strategies, we categorized the posts into five predefined topics—engagement, research, image, society, and education. This categorization, combined with Long Short-Term Memory (LSTM) and a Random Forest (RF) algorithm, was utilized to predict social media output in the last five days of each month, achieving successful results. This paper further explores how variations in these social media strategies correlate with the rankings of HEIs. Our findings suggest a nuanced interaction between social media engagement and the perceived prestige of HEIs.

Abstract
Automated assessment tools for programming assignments have become increasingly popular in computing education. These tools offer a cost-effective and highly available way to provide timely and consistent feedback to students. However, when evaluating a logically incorrect source code, there are some reasonable concerns about the formative gap in the feedback generated by such tools compared to that of human teaching assistants. A teaching assistant either pinpoints logical errors, describes how the program fails to perform the proposed task, or suggests possible ways to fix mistakes without revealing the correct code. On the other hand, automated assessment tools typically return a measure of the program's correctness, possibly backed by failing test cases and, only in a few cases, fixes to the program. In this paper, we introduce a tool, AsanasAssist, to generate formative feedback messages to students to repair functionality mistakes in the submitted source code based on the most similar algorithmic strategy solution. These suggestions are delivered with incremental levels of detail according to the student's needs, from identifying the block containing the error to displaying the correct source code. Furthermore, we evaluate how well the automatically generated messages provided by AsanasAssist match those provided by a human teaching assistant. The results demonstrate that the tool achieves feedback comparable to that of a human grader while being able to provide it just in time.

Abstract
As social media (SM) becomes increasingly prevalent, its impact on society is expected to grow accordingly. While SM has brought positive transformations, it has also amplified pre-existing issues such as misinformation, echo chambers, manipulation, and propaganda. A thorough comprehension of this impact, aided by state-of-the-art analytical tools and by an awareness of societal biases and complexities, enables us to anticipate and mitigate the potential negative effects. One such tool is BERTopic, a novel deep-learning algorithm developed for Topic Mining, which has been shown to offer significant advantages over traditional methods like Latent Dirichlet Allocation (LDA), particularly in terms of its high modularity, which allows for extensive personalization at each stage of the topic modeling process. In this study, we hypothesize that BERTopic, when optimized for Twitter data, can provide a more coherent and stable topic modeling. We began by conducting a review of the literature on topic-mining approaches for short-text data. Using this knowledge, we explored the potential for optimizing BERTopic and analyzed its effectiveness. Our focus was on Twitter data spanning the two years of the 117th US Congress. We evaluated BERTopic's performance using coherence, perplexity, diversity, and stability scores, finding significant improvements over traditional methods and the default parameters for this tool. We discovered that improvements are possible in BERTopic's coherence and stability. We also identified the major topics of this Congress, which include abortion, student debt, and Judge Ketanji Brown Jackson. Additionally, we describe a simple application we developed for a better visualization of Congress topics.

Abstract
Static source code analysis techniques are gaining relevance in automated assessment of programming assignments as they can provide less rigorous evaluation and more comprehensive and formative feedback. These techniques focus on source code aspects rather than requiring effective code execution. To this end, syntactic and semantic information encoded in textual data is typically represented internally as graphs, after parsing and other preprocessing stages. Static automated assessment techniques, therefore, draw inferences from intermediate representations to determine the correctness of a solution and derive feedback. Consequently, achieving the most effective semantic graph representation of source code for the specific task is critical, impacting both techniques' accuracy, outcome, and execution time. This paper aims to provide a thorough comparison of the most widespread semantic graph representations for the automated assessment of programming assignments, including usage examples, facets, and costs for each of these representations. A benchmark has been conducted to assess their cost using the Abstract Syntax Tree (AST) as a baseline. The results demonstrate that the Code Property Graph (CPG) is the most feature -rich representation, but also the largest and most space -consuming (about 33% more than AST).