FewShot AdaptiveMLC is a two-phase learning framework that first applies multi-task learning (MTL) to build a shared cross-label representation using XLM-RoBERTa, followed by a targeted few-shot fine-tuning (FSFT) stage tailored to minority labels. The second stage freezes the shared encoder and updates only the task-specific heads using small, balanced few-shot subsets, enabling focused adaptation without catastrophic forgetting. Experiments demonstrate substantial improvements over standard MTL and fine-tuning baselines on the persistent challenge of class imbalance and the poor performance of minority labels that traditional oversampling or augmentation techniques cannot resolve. Our approach achieves an absolute +0.072 gain in macro-F1 (from 0.712 to 0.784), including a +0.183 increase in F1 for Disability (from 0.506 to 0.689). These results show that targeted adaptation effectively recovers underrepresented labels while maintaining stable performance in majority classes. Unlike prior multi-task or data augmentation approaches that treat all labels uniformly, the novelty of this work lies in decoupling shared representation learning from minority-label specialization through a dedicated, gradient-isolated few-shot stage, offering a practical and scalable solution for multi-label hate speech detection in low-resource NLP settings. BOISHOMMO, a new manually annotated multi-label Bangla hate speech dataset, was introduced and used to validate this approach.
Prediction of stock prices ranks among the most difficult problems in quantitative finance because of dynamic, nonlinear, and emotional nature of financial markets. Learning buy–hold–sell strategy entails both learning from past price patterns and understanding the investor psychology and reacting to market dynamics rapidly. The problem of stock market decision-making consists not only in predicting the future prices but also in allocating investments dynamically in uncertain environment with transaction costs. The present research aims to investigate the problem of sentiment-based portfolio optimization, which is represented as a problem of reinforcement learning that derives trading strategies based on financial and market sentiment indicators. While machine learning techniques such as long short-term memory (LSTM) and BiLSTM with attention are known for being able to recognize temporal relations and sequences of events, they lack the ability to consider psychological and event-based dynamics of financial markets and are unable to detect rapid changes in sentiments. Although RL approaches such as deep Q-networks (DQN) and proximal policy optimization (PPO) are efficient in terms of optimization of actions, they suffer from poor reward propagation and poor exploration in unstable environment and cannot use different sources of information. In order to address this challenge, a new hybrid framework has been proposed which exploits a novel DDQN framework, considering the signals that arise out of the sentiments of the market created through news and social media as well as the signals coming from financial signals. This hybrid framework is capable of exploiting the sentiment analysis framework in order to analyze the sentiments and volatilities of the investors, whereas the engineered financial signals would generate price, volume, and momentum information. As one can notice from the experiment results, the hybrid model which has been proposed outperforms all the other considered models in terms of such performance metrics as accuracy (0.9556), precision (0.9605), recall (0.9556), and F1-score (0.9558) in comparison with BiLSTM + attention (accuracy 0.795, F1 0.796), PPO (0.78, 0.781), LSTM (0.768, 0.767), and DQN (0.755, 0.754). Besides, as one can observe from the confusion matrix analysis, the least number of mistakes is made when classifying hold, invest, and sell actions. As one can see from the ROC curves, the hold, invest, and sell actions are most separable since their AUC is equal to 0.94, 0.92, and 0.90, correspondingly. Moreover, the graph of the reward per stock demonstrates the flexibility of the algorithm concerning stable and volatile stocks and earning of enough cumulative rewards that make up more than 3.0–6.0 Sharpe ratios.
Financial fraud detection presents significant challenges, including mixed data types, extreme class imbalance, and the need for model interpretability. This paper introduces INSIGHT, a novel supervised cluster-based framework that advances data science methodologies for fraud detection. Our approach integrates a computationally efficient mixed-distance metric for handling continuous and categorical features, together with cluster-aware oversampling that preserves data distribution integrity. The framework establishes theoretical guarantees for algorithmic convergence, cluster stability, and scalable performance. Comprehensive evaluation on two financial datasets, European Credit Card (284,807 transactions) and BankSim (594,643 transactions), demonstrates state-of-the-art performance, with F1 scores of 0.82 and 0.71: INSIGHT attains the highest F1 and AUC of all compared methods on the European Credit Card dataset, ahead of tuned CatBoost and a panel of supervised base models, and the highest AUC and precision on BankSim, while outperforming isolation forests and autoencoder-based anomaly detectors on both datasets. Ablation studies isolate the individual contribution of the mixed-distance metric, cluster-aware oversampling, and the choice of k-medoids variant, while targeted experiments connect the theoretical guarantees to measured robustness and scalability. INSIGHT’s inherent interpretability provides transparent cluster characterization and feature importance analysis, offering actionable insights for financial security applications. The framework combines theoretical soundness with practical scalability for financial analytics.
We introduce OFTER (online forecasting via temporal embedding regression), a time series forecasting pipeline tailored for financial problems centered on mid-sized multivariate time series. OFTER utilizes the non-parametric models of k-nearest neighbors and generalized regression neural networks, integrated with a dimensionality reduction component. To circumvent the curse of dimensionality, we employ a weighted norm based on a modified version of the maximal correlation coefficient. The pipeline we introduce is specifically designed for online tasks, has an interpretable output, and is able to outperform several state-of-the-art baselines. The computational efficacy of the algorithm, its online nature, and its ability to operate in low signal-to-noise regimes, render OFTER an ideal approach for financial multivariate time series problems, such as daily equity forecasting. Our work demonstrates that while deep learning models hold significant promise for time series forecasting, traditional methods carefully integrating mainstream tools remain very competitive alternatives with the added benefits of scalability and interpretability.
Because of the rapid development of large language models (LLMs), distinguishing AI-generated text from human-written text has become an important issue for ensuring validity, reliability, and accountability in digital communication. This research presents a fuzzy logic-based framework for differentiating AI-generated from human-authored textual content by modeling linguistic uncertainty and reasoning based on if–then rules. Six stylistic and linguistic parameters such as lexical variability, content word ratio, lexical density, keyword density, Coleman–Liau readability index, and Flesch reading ease score are utilized as input and mapped by using triangular membership functions. The proposed model computes a human-likeness score (HL score) and provides subtle insight into textual authenticity. The model is evaluated on five domain-diverse datasets—Play Store app reviews dataset, financial dataset, Twitter sentiment dataset, IMDB reviews dataset, and social media dataset; each holds 500 sentences. The presented model achieves 100
