Temporal reasoning in financial texts is essential for understanding event timing and claim validity, especially in earnings conference calls and social media discussions. While transformer-based models have advanced natural language processing, the comparative performance of fine-tuned encoder models and prompt-based decoder models in multilingual temporal classification remains underexplored. This study systematically compares model types, model sizes, and prompting strategies across two tasks: detecting temporal references in English texts and assessing claim validity in Chinese posts. Encoder models such as RoBERTa and BERT and decoder models such as GPT-4o, Mistral, and Gemma are evaluated using fine-tuning and few-shot prompting approaches. Results show that fine-tuned encoder models achieve consistently strong performance across both English and Chinese datasets. Mid-sized prompt-based decoder models also perform competitively under well-designed prompts, offering a practical alternative when fine-tuning is not feasible. In addition, decoder models are more robust to class imbalance, as reflected by smaller gaps between Micro-F1 and Macro-F1 scores. However, decoder models perform less effectively on Chinese tasks, indicating the need for language-specific adaptation. These findings provide practical guidance for selecting models and designing prompts for financial natural language processing under resource constraints.
Large Language Models (LLMs) exhibit impressive capabilities in natural language understanding and generation; however, their ability to interpret and integrate nonverbal emotional cues, such as facial expressions, while maintaining security and interpretability remains underexplored. This study investigates how multimodal LLMs, specifically Qwen2.5-VL and Deepseek-VL, respond to conversational prompts paired with facial expression images through the lens of neural-symbolic integration. We constructed a dataset of 10,000 conversational lines combined with real and synthetic facial expressions depicting various emotional states. Using both automated sentiment analysis and human evaluations based on a 5-point Likert scale, we assessed model responses for tone appropriateness, helpfulness, and emotional alignment. Our results indicate that neural-symbolic integration significantly enhances interpretability and robustness against adversarial inputs, enabling models to achieve a higher average similarity with human interpretations (cosine similarity peaking around 0.7–0.9) compared to Qwen2.5-VL (0.45–0.55). However, both models struggled to accurately interpret subtle or mismatched emotional cues. These findings highlight the potential of neural-symbolic integration to improve the security and emotional reasoning of AI systems.
This study examines whether targeted feedback can lead to improvements in nonverbal communication, specifically in speech delivery, by analyzing 3-min videos of three management consultants. The analysis focused on facial expressions, utilizing the Facial Action Coding System (FACS) and vocal characteristics, as represented by Mel-Frequency Cepstral Coefficients (MFCCs). For five months, the authors provided individualized feedback focusing on facial expression and speech delivery, which involves key elements such as speaking pace, intonation, and pitch variation. While the sample size was limited to three individuals, the study leveraged a year’s worth of accumulated video data. The findings indicate that feedback contributed to observable improvements in facial expressions, whereas vocal features derived from MFCCs remained largely unaffected. Furthermore, this study suggests that combining FACS with real-time speech content extraction offers a deeper understanding of how message content and emotional expression interact, and using the T5 model to generate feedback on speech content automatically improves the speech.
Social media platforms like X (formerly Twitter) play a central role in public discourse but are also exploited for influence operations (IO) through coordinated inauthentic behavior (CIB). This study proposes a method to detect IO-related coordinated communities during the August 2023 release of Advanced Liquid Processing System (ALPS)-treated water from the Fukushima Daiichi Nuclear Power Plant. Using reposting data, we construct a graph of user communities based on network science techniques, incorporating both intra- and inter-community features. A Graph Neural Networks (GNN) is trained on these structures to classify communities as abnormal or normal. The model achieves F1 = 0.97, outperforming baseline methods. By automating early detection of coordinated communities, our method supports timely countermeasures in IO. While effective, further work is needed to capture communities with varied intents and improve attribution. The results demonstrate the potential of graph-based learning in real-time monitoring of influence activities on social platforms.
Managing the privacy of social media posts remains a complex task, especially as audience diversity and content sensitivity grow. We propose a comprehensive privacy management framework that combines post content features with behavioral signals from social interactions to deliver personalized audience recommendations. Leveraging Facebook and Reddit datasets, we implement five core modules: post privacy classification, persona contradiction detection, interaction and privacy alignment scoring, expectation mismatch analysis, and privacy-aware friend grouping. Our post classifier achieves F1-scores of 0.76 (Facebook) and 0.72 (Reddit); contradiction detection yields an F1-score of 0.80 by combining behavioral and BERT-based features. Friend clustering based on interaction and alignment scores results in silhouette scores of 0.65 (Facebook) and 0.60 (Reddit), while expectation mismatch analysis reveals stronger emotional alignment in highly private posts. Overall, our approach enables explainable and behavior-sensitive audience control for social platforms, improving upon prior content- or interaction-only methods.
In this study, we evaluate and compare two feature optimization methods for enhancing Intrusion Detection Systems (IDS): Genetic Algorithm (GA)-based and Decision Tree (DT)-based approaches. Both methods aim to select optimal feature subsets to improve classification performance and reduce computational cost. The GA-based approach integrates a novel fitness function with Least Squares Support Vector Machine (LSSVM) to simultaneously maximize the True Positive Rate (TPR), minimize the False Positive Rate (FPR), and optimize features. The DT-based method leverages a hybrid filter-wrapper strategy using C5.0 decision trees and LSSVM, employing gain ratio and pruning for initial selection, followed by predictor importance ranking for refined optimization. Experimental evaluations using the KDD CUP 99 and UNSW-NB15 datasets demonstrate that the DT-based feature selection consistently outperforms the GA-based method across most attack categories in terms of accuracy, TPR, FPR, and Area Under the Receiver Operating Characteristic (ROC) Curve (AUC). The results show that DT has better performance than GA with respect to feature optimization.
In this paper we present a comprehensive study on hate speech detection in Bengali (Bangla), a low-resource language with significant online presence. We explore the potential of large language models (LLMs) such as GPT-4, Qwen, and DeepSeek in identifying hate speech from social media content, including transliterated and code-mixed text. Using a consolidated dataset combining multiple public hate speech corpora, we evaluate LLM-based prompting and fine-tuning strategies alongside traditional deep learning and transformer models. Our findings show that fine-tuned LLMs like DeepSeek-67B and GPT-4 consistently outperform smaller models, achieving macro-F1 scores above 90
Scientific research is expanding rapidly, producing vast volumes of scholarly literature across diverse disciplines. However, identifying emerging trends and uncovering interdisciplinary intersections remains a persistent challenge due to the fragmented and nonlinear nature of modern knowledge production. Traditional bibliometric methods such as co-citation analysis and keyword co-occurrence rely on pairwise relationships and often fail to capture the higher-order associations that drive innovation. In this study, we present a novel hypergraph-based framework for forecasting scientific research trends. In our approach, nodes represent research concepts, and hyperedges correspond to publications that link multiple concepts. By framing the task as a hyperedge link prediction problem, we uncover latent conceptual groupings that may signal future research directions. We then develop hypergraph neural network models, HLP, HyperGCN, and HyperSAGE, alongside traditional graph-based models, using a real-world dataset of scientific concepts used in papers from the field of Ethnic Studies. Our results show that hypergraph models consistently outperform their graph-based counterparts in accuracy and predictive power.
Personalized recommendation vectors are often correlated with the global feature vector, such as a popularity vector. Prior works tried to adjust the correlation using parameters. However, there are no parameter setting guidelines, and it is unclear whether the desired correlation is achieved. We propose CoCoA to generate the recommendation vector with a linear combination of the given source and global vectors. It ensures the cosine similarity between the recommendation and global vectors is a user-input value. A case study with a movie dataset observed CoCoA can flexibly control the effect of popular movies. Evaluations on three real-world graph datasets showed that only CoCoA achieved negative cosine similarity, where the globally important nodes are suppressed.
This study analyzes propaganda from the Terrorgram Collective, a decentralized far-right network of Telegram channels and users that promote violent accelerationism. Drawing from three key Terrorgram publications disseminated between 2021 and 2023, 393 instances of threatening communication have been extracted and analyzed. The threats are categorized along four analytical dimensions: type (general violence vs. specific plans), mode of communication (direct, indirect, veiled), target type (soft vs. hard), and target group. The analysis suggests that threats in Terrorgram publications are mainly directed at soft targets—i.e., individuals and minority communities—and primarily communicated through veiled or indirect language. Moreover, threatening communication in Terrorgram publications frequently includes operational guidance designed to encourage lone-actor violence.
The rise of do-it-yourself (DIY) cosmetic procedures promoted on social media platforms such as TikTok has introduced new risks to public health, particularly as untrained individuals attempt at-home dermal filler injections. This study investigates the feasibility of detecting medical misinformation related to DIY fillers using machine learning techniques and examines the impact of annotation consistency on model performance. We collected and manually labeled 195 TikTok posts using a three-class schema: non-relevant, relevant-benign, and relevant-misinformation. Labels were assigned by both a medical expert and a technical contributor, with a subset re-labeled to assess intra-annotator agreement. Results showed moderate-to-substantial agreement within the same expert (κ = 0.624) but low agreement across annotators, revealing variability in label interpretation. A Random Forest classifier trained on different label subsets showed that annotations from the more internally consistent rater led to stronger model performance, particularly in precision. These findings underscore the importance of early investment in annotation quality and inter-rater validation when building AI systems for misinformation detection. We discuss the implications for public health surveillance and propose future work to scale content filtering and support qualitative review by experts.
This study developed an automated system leveraging the Llama 3.1 8B large language model, specifically designed for generating sustainability reports. The results indicate that, while the model performed well on the training data—demonstrating high accuracy and low loss—its generalization capabilities still require enhancement. This research not only demonstrates the feasibility of using large language models to automatically produce professional sustainability reports but also provides valuable methodological insights and a technical foundation for future studies in related fields.
Online transactions platforms such as e-Commerce and social networks generate massive volumes of data, which are complex and with sizes that can reach terabytes ( 10^12 bytes) or petabytes ( 10^15 ), making data mining and recommendation tasks increasingly challenging. Existing systems for high utility sequential pattern mining (HUSPM) extract valuable (e.g., profitable) e-commerce sequential products to recommend to buyers considering both frequency of items and utility (e.g., profit, importance) of itemsets. Existing HUSPM systems include those named as HUSREC21, HUSP21, and HUSP-SP23, mine high utility sequential patterns using a single machine for these HUSPM tasks, resulting in inefficient processing of big sized datasets, longer execution times and high memory consumption. This paper proposes a system called Big High Utility Sequential Pattern Recommendation System (BIGHUSREC), that extends the HUSREC21 system to mine high-utility sequential patterns from big sized datasets through a "Top-K" approach integrated with the MapReduce framework. The focus is on extracting the Top-K most valuable (profitable) and yet relevant to the user patterns, with the aim of improving recommendation accuracy, while minimizing execution time. The proposed system uses MapReduce to partition data into smaller parts (Mapping), and analyzing each part in parallel to identify profitable patterns before aggregating the results (Reducing) for a comprehensive output. By combining purchase and clickstream data, the proposed BIGHUSREC effectively improves recommendation accuracy.
As large language models (LLMs) increasingly engage in emotionally sensitive interactions, their unintended therapeutic roles demand systematic investigation. We examine ChatGPT’s perceived capacity to provide emotional and mental health support by analyzing user-generated content from social networks through relevance classification and sentiment analysis. Guided by three research questions, we quantify public sentiment toward ChatGPT in therapeutic contexts. To identify posts suggesting therapeutic use of ChatGPT, we introduce two methods: SemReC, a supervised relevance classification, and PASS, an unsupervised similarity-based approach. Both methods demonstrate high accuracy and consistent performance across the dataset. We further assess the performance of existing pre-trained sentiment analysis models to benchmark their effectiveness. To capture affective sentiment propagation in multi-turn interactions, we propose two tree-structured methods—HierSent and AggSent—which model emotional dynamics within threaded conversations. Empirical results validate the effectiveness of our methods and reveal a predominance of positive sentiment toward using ChatGPT for therapeutic purposes. These findings highlight the public popularity of the emergent therapeutic use of LLMs and underscore the need to examine their broader implications for mental health.
Recurrent waves of infection are influenced by a combination of factors, including seasonal effects, the emergence of new pathogen strains, and fluctuations in human behaviour. Social networks play a key role in shaping individual perceptions of infection risk. However, capturing the complex influence of social media on individual attitudes toward public health measures—such as social distancing—remains a significant modelling challenge. In this study, we investigate how social network topology impacts opinion dynamics related to risk during a pandemic. Building on a previously calibrated and validated agent-based model of the COVID-19 pandemic, we incorporated a social network layer to simulate the spread of risk perception. This enabled us to compare the effects of different network structures on the emergence of recurrent waves. Our simulations indicate that networks exhibiting both scale-free and small-world properties most closely reproduce real-world infection dynamics.
The widespread deployment of intelligent systems in critical sectors, such as healthcare, justice, and finance has underscored the urgent need for transparent and trustworthy in the answers provided. Accordingly, explainability constitutes a key requirement to foster understanding, trust, and accountability in automated decision-making processes where AI in general, and Large Language Models (LLMs) in particular, are involved. This paper introduces a new explainability framework for symbolic systems, centered on an application pipeline aimed at constructing a semantic knowledge graph from unstructured text. In our approach, explainability for the symbolic entity recognition component is achieved through an event-driven system, where each operation is logged atomically in a journal, capturing events that trace object creation, lineage, and purpose. Alongside the journal, two additional components support the architecture: the Entity Storage, which maintains versioned records of entities, and the Explainability Module, which interprets and explains the logic behind each entity's representation. The proposed framework offers a high degree of detail and customizability and is currently being tested within a medical application designed for general practitioners. In this context, the ability to provide transparent justifications for specific recommendations is essential to ensuring reliability for both physicians and patients. Furthermore, it contributes to strengthening the credibility and trustworthiness of AI-based tools, which are expected to assume an increasingly central role in this domain.
The development of antipsychotic drugs poses challenges due to the complex pharmacology of psychiatric drugs and the limited diversity of existing drugs. Recent advances in deep-generative models have enabled the exploration of vast chemical spaces for the design of novel drugs. In this study, we leverage the use of the Llama 3.2 1B model to predict potential antipsychotic drugs. We use the ligands and target information in the BindingDB dataset. The ligands were represented using the SELFIES and Group SELFIES notation. Initially, the model was trained for unconditional generation. We also studied the effect of relative attention on the model performance using both the SELFIES and Group SELFIES representation. Then we use a pre-trained model to fine-tune the ligand-target dataset in order to discover molecules that bind to Dopamine D2 and Serotonin receptors (the primary receptors for antipsychotics). We found 10 potential candidates. We then screen them for drug candidacy. Next, a docking analysis was conducted on the screened drugs to find the binding affinity of the drug to the different proteins within the receptors. Finally, we suggest 4 molecules as potential antipsychotic drugs.
Predicting individual player performance, particularly scoring metrics like points per game (PPG), has become a significant area of research in sports analytics, driven by advances in data collection and machine learning. Previous studies primarily emphasized team performance or used aggregated data without in-depth feature optimization, leaving gaps in accurately forecasting individual player metrics. This paper addresses these limitations by introducing the Correlation-Optimized NBA Scoring Estimator (CONSE) model, designed to predict NBA player performance based on statistical data spanning ten NBA seasons (2013–2023). Our approach includes rigorous data preprocessing, extensive exploratory data analysis (EDA), and leverages correlation-based feature selection to enhance interpretability and predictive accuracy. Multiple regression models, including Multiple Linear Regression, K-Nearest Neighbors (KNN), Decision Tree, and Random Forest, were evaluated within the CONSE model framework using R-squared and Mean Absolute Error along with cross-validation. Random Forest Regression demonstrated superior performance due to its robustness against overfitting and outliers. Our findings provide valuable insights for coaches, analysts, and enthusiasts aiming to identify future NBA stars using a data-driven approach.
Deep Learning is a new branch of Machine Learning that focuses on applying Artificial Neural Networks to obtain more cluttered decision boundaries. However, the supervised Neural Networks are sensitive to presentation order, architecture configuration, and complex shapes. In addition, the learning instability causes large changes in its performance on training samples. In our study, we proposed a new model for RBF Neural Network tuning by using fuzzy consensus clustering and model selection. Experimental studies showed that our Deep Learning model, which is named DeepRBF, yielded better recognition accuracy than other supervised learning algorithms. In addition, our typical initialization scheme speeds up the learning convergence and avoids the local minimum.
This paper presents the design and manufacturing of a personalized ankle prosthesis tailored to meet the diverse treatment needs of individuals with ankle disorders. The proposed approach integrates advanced 3D modeling and printing techniques to develop custom-made implants that optimize mobility, functionality, and patient comfort. By leveraging 3D printing technology, biocompatible and durable materials are utilized to create implants that closely match the unique anatomical structure of each patient’s joint. This personalized approach aims to enhance orthopedic treatments by reducing anesthesia time, minimizing post-operative complications, and improving overall patient care. The study highlights the potential of personalized prosthetic solutions in advancing orthopedic medicine and improving the quality of life for affected individuals.
