The spread of toxic and profane language across online platforms has necessitated the development of advanced automated detection systems to maintain safe online environments. This survey comprehensively explores the progression of profanity language detection techniques, focusing on datasets, feature extraction methods, and model architectures. Through a thorough examination of key developments in the field, the study outlines the evolution of detection systems across four distinct phases. It offers a taxonomy of approaches that have shaped the current landscape. We review various high-performing models, highlighting their effectiveness in addressing challenges such as context understanding, multilingual data, and identifying subtle forms of toxicity. The article further explores the strengths of these models, including their ability to capture long-range dependencies and nuanced features, while also discussing their limitations, such as computational efficiency and generalization across datasets. By analyzing the current state of toxic language detection and the latest advancements in feature extraction and modeling techniques, we provide valuable insights for researchers and practitioners. Additionally, the article outlines future research directions, emphasizing the importance of lightweight, adaptive models and real-time optimization techniques to enable scalable, practical deployment across diverse platforms and environments.
Forecasting unemployment rates using multivariate time series is a tough problem due to its complex temporal dependencies, demographic complexity and need for uncertainty quantification that is interpretable. In labour economics and macro-econometrics, the demographic complexity of the unemployment rate dynamics has been well studied; the recent approaches for deep learning forecasting barely take into account the demographic complexity of the unemployment dynamics. We present SEMINAR: sparsity-enhanced Mamba (SEM), economic priors integration layer (EPIL), adaptive cross-variable attention (ACVA), seasonally-naive anchor-based iterative refinement regression block (IRRB) and probabilistic uncertainty quantifier (PUQ), giving heteroscedastic, horizon-aware confidence intervals. This proposed model is tested against autoregressive integrated moving average (ARIMA), long short-term memory (LSTM), non-sparse Mamba, and transformer in three progressively harder settings using international data for unemployment rate forecasts (2006–2024) in a leak-free chronological protocol with 12-step-ahead multi-horizon forecasting and statistically significant performance evaluation across eight random seeds. In the base setting, SEMINAR gives the best mean absolute error of all five models (128.6 vs . 130.2–149.0 in baselines), beating ARIMA ( p = 0.008) and transformer ( p = 0.039). Under cross-country transfer, SEMINAR is the strongest neural model on the Korea/Chile/Russia task mean absolute error (MAE) (134.6 MAE), although the classical ARIMA baseline remains substantially more accurate there (86.0 MAE); on the short, weakly seasonal Bolivia/Dominican Republic task SEMINAR is outperformed by all four baselines, a failure mode we trace to the seasonal-anchor assumption. Both boundaries are reported rather than omitted. SEMINAR also gives calibrated 80% prediction intervals, a feature that none of the baselines provides in their native form. The complete architecture consists of around 44,401 parameters; an ablation study with seven variants has been done.
Psychological AI advocates designing algorithms inspired by human cognition. We apply this paradigm to Entity Matching (EM), where high-stakes scenarios demand not only high accuracy but also human-like reasoning to ensure transparency, reliability, and user confidence. However, current deep learning (DL) EM models often act as black boxes, failing to capture human decision-making processes, such as prioritizing diagnostic attributes, inferring missing information from context, and assessing overall semantic consistency. We propose cueEM, a human-like EM model grounded in Cue Validity Theory. By adopting a text-centric design and leveraging Bert-Pair-Networks with a hybrid attention mechanism, cueEM captures and integrates textual features across attributes. It generates attribute-specific weights that explicitly reveal each attribute’s contribution, aligning the model’s decisions with human decision processes. Extensive evaluations on DeepMatcher benchmarks demonstrate that cueEM outperforms Effective Entity Matching with Transformers (DITTO) and achieves competitive performance with large language models (LLMs), while providing inherently human-like reasoning. To our knowledge, cueEM is the first text-focused framework that combines hybrid attention with human-aligned reasoning, enabling interpretable and human-like decision-making.
Timely intervention is critical for saving a patient’s life. To monitor vital signs, various health monitoring sensors are placed on the patient’s body, collecting data at regular intervals. However, the limited computational capacity of these health monitoring sensor nodes prevents them from analyzing the data independently. By integrating sensors with Internet of Things (IoT) nodes, the collected data can be offloaded to nearby fog nodes for processing. A challenge arises when certain fog nodes receive a disproportionate number of healthcare data tasks compared to others, necessitating an efficient load balancing mechanism to process these time-sensitive tasks promptly. In this work, we propose an Intelligent Task Offloading Scheme for Healthcare Data Analysis ( ITOS HDA ). The scheme consists of two key algorithms: (1) a load balancing algorithm that employs the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) method to allocate offloaded tasks efficiently among fog nodes, and (2) a task execution algorithm that uses the 0/1 Knapsack approach to prioritize tasks based on their deadlines. The performance of ITOS HDA is evaluated against state-of-the-art schemes, demonstrating its superiority in handling varying numbers of fog nodes and offloaded tasks. Experimental results confirm that ITOS HDA outperforms existing algorithms in terms of efficiency and scalability.
Accurate prediction of student academic performance is crucial for enhancing educational outcomes and enabling personalized learning interventions. However, the limited availability of labeled data in many practical educational settings poses significant challenges for conventional machine learning models, which often require large datasets to achieve satisfactory performance and are prone to overfitting. To address this issue, we propose a novel small-sample learning framework based on the Tabular Prior-Data Fitted Network (TabPFN). By integrating a Transformer based architecture with Bayesian inference, TabPFN effectively leverages meta-learned priors to perform high-quality predictions without task-specific finetuning. We evaluated the approach on the Mathematics and Portuguese subsets of the UCI Student Performance dataset and on a pooled Combined dataset under small-sample conditions. Experimental results indicate that TabPFN achieved generally strong and competitive performance across the three datasets, with accuracies of 92.36%, 95.49%, and 94.29%, respectively. However, its advantage varied across datasets and metrics, and the results should be interpreted cautiously where confidence intervals overlapped with strong baseline models. Additionally, we employ Shapley Additive Explanations (SHAP) to interpret model decisions, identifying key influencing factors such as prior grades (G1, G2) and prior educational experience. The findings indicate that TabPFN constitutes a promising and effective solution for educational prediction in data scarce scenarios but also provides actionable insights for instructional design and student support.
