Credit card fraud detection remains a critical challenge in financial security, characterized by evolving fraud patterns, sparse labeled data, and severe class imbalance. Traditional methods often fail to capture the complex temporal dynamics and heterogeneous relationships inherent in financial transaction networks. To address these limitations, we propose TH-GCL (Temporal Heterogeneous Graph Contrastive Learning), a novel framework that integrates heterogeneous graph modeling, temporal pattern recognition, and contrastive learning for enhanced fraud detection. Our approach constructs a temporal heterogeneous graph incorporating multiple entity types including users, transactions, merchants, and devices, with time-aware edge weights to capture evolving behavioral patterns. We design a temporal-aware graph neural network architecture that learns hierarchical representations by jointly modeling structural dependencies and temporal evolution patterns. Furthermore, we introduce a dual-view contrastive learning mechanism that creates augmented graph views through both structural perturbation and temporal masking, enabling the model to learn robust representations under different perspectives. The contrastive objective encourages the model to distinguish between normal and fraudulent transaction patterns while maintaining consistency across augmented views. Extensive experiments on the IEEE-CIS Fraud Detection dataset demonstrate that TH-GCL achieves superior performance compared to state-of-the-art baselines, with improvements of 5.2% in AUC-ROC and 8.7% in AUC-PR. Ablation studies confirm the effectiveness of each component, while analysis of learned representations reveals meaningful fraud pattern discovery. Our framework exhibits strong generalization capability across different time periods and transaction volumes, making it practical for real-world deployment in dynamic fraud detection scenarios.

Cryptocurrency markets exhibit complex microstructural dynamics characterized by high-frequency volatility bursts, rapid regime switching, and long-range temporal dependencies, which expose several limitations of existing volatility forecasting approaches. In particular, attention-based models suffer from prohibitive quadratic computational cost on long high-frequency sequences, while many recurrent architectures struggle to adapt to regime transitions, asymmetric volatility responses, and risk-aware uncertainty estimation. To address these gaps, this paper proposes CryptoMamba-SSM, a novel volatility prediction framework built upon Mamba-based state space models with linear computational complexity. CryptoMamba-SSM integrates selective memory mechanisms with structured state space representations to effectively capture critical market microstructure signals arising from liquidity shocks and sentiment transitions, while dynamically adjusting memory retention across different volatility regimes. This design enables efficient modeling of long-sequence dependencies inherent in cryptocurrency price movements without incurring the computational bottlenecks of traditional attention-based architectures. Through comprehensive experiments on Bitcoin historical data spanning multiple market regimes, we demonstrate that CryptoMamba-SSM consistently outperforms conventional LSTM, GRU, and Transformer baselines, achieving up to a 23.7% reduction in Mean Absolute Error and a 31.2% improvement in directional accuracy. The selective memory mechanism effectively captures regime-switching behaviors and microstructural anomalies, leading to more reliable short-term volatility risk quantification. Moreover, the linear-time complexity of CryptoMamba-SSM enables real-time processing of high-frequency trading data while maintaining strong generalization across diverse market conditions.

Multi-agent reinforcement learning for supply chain optimization faces significant challenges in maintaining physical consistency, quantifying operational risks, and coordinating distributed decision-making. Existing approaches frequently violate fundamental conservation principles, rely on expected value optimization that ignores tail risks, and lack effective mechanisms for ensuring global coherence across autonomous agents. We present PIMA-DRL, a unified framework that combines physics-informed neural networks with distributional reinforcement learning to simultaneously address these limitations. The framework enforces conservation laws through differentiable physics constraints embedded directly into the learning process, while maintaining complete probability distributions over returns to enable sophisticated risk assessment through Conditional Value-at-Risk optimization. A novel coordination mechanism based on Lagrangian duality ensures global consistency while preserving agent autonomy through decentralized multiplier updates. We establish theoretical convergence guarantees under bounded physics constraint violations and prove enhanced sample efficiency compared to physics-agnostic methods. Local agent states encompass inventory levels, flow dynamics, demand projections, and capacity limitations, with evolution governed by conservation-based differential equations. Comprehensive experiments across multi-echelon inventory networks, hub-spoke distribution systems, and perishable goods supply chains demonstrate substantial improvements over conventional baselines. Results show conservation error reductions of 87.3%, constraint violation decreases of 74.6%, and tail risk improvements of 42.1% while achieving 2.4x faster convergence. The physics-informed structure enhances model interpretability and provides principled uncertainty quantification critical for robust supply chain management under operational variability.

The accurate modeling of implied volatility surfaces is crucial for option pricing and risk management in financial markets. Traditional parametric approaches, such as the Stochastic Volatility Inspired (SVI) model, often suffer from rigid functional forms that inadequately capture the complex nonlinear patterns observed in real market data, particularly in the tail regions of volatility smiles. This paper introduces a novel Generative Adversarial Network (GAN) framework specifically designed to learn and reconstruct implied volatility surfaces directly from historical option market prices. Our approach leverages the adversarial training mechanism to generate smooth, market-consistent volatility surfaces that overcome the limitations of conventional parametric models. The proposed GAN architecture incorporates domain-specific constraints and regularization techniques to ensure the generated surfaces satisfy fundamental no-arbitrage conditions while maintaining computational efficiency. Through comprehensive empirical evaluation using real-world AAPL options data spanning 2016-2020, we demonstrate that our GAN-based approach significantly outperforms traditional methods in reducing pricing errors under the Black-Scholes framework. Specifically, our method achieves a 23.7% reduction in mean absolute percentage error (MAPE) for out-of-the-money options and a 31.4% improvement in root mean square error (RMSE) compared to the baseline SVI model. The results indicate particularly substantial improvements in pricing accuracy for deep out-of-the-money options, where traditional models typically exhibit the largest deviations. Furthermore, our approach demonstrates superior stability across different market regimes and provides enhanced generalization capabilities for options with varying maturities and moneyness levels. The proposed methodology offers significant practical implications for derivatives trading, risk management, and model validation in quantitative finance.

Credit card fraud detection remains a critical challenge in financial security, characterized by severe class imbalance and the need to capture complex temporal patterns in transaction sequences. Traditional machine learning approaches treat transactions as independent events, failing to model the sequential nature of user behavior and suffering from inadequate handling of minority class samples. In this paper, we propose an integrated framework that combines generative modeling and time-aware sequential learning for credit card fraud detection. Our approach addresses two fundamental limitations: (1) we model transaction histories as temporal sequences using a Transformer-based architecture that captures both long-term dependencies and abrupt behavioral changes through multi-head self-attention mechanisms, and (2) we employ CTGAN to generate high-quality synthetic fraudulent samples, providing more effective oversampling than conventional techniques like SMOTE. The Time-Aware Transformer incorporates temporal encoding and position-aware attention to preserve transaction order and time intervals, while CTGAN learns the complex conditional distributions of fraudulent transactions to produce realistic synthetic samples. We evaluate our framework on the IEEE-CIS Fraud Detection dataset, demonstrating significant improvements over representative classical and sequential deep-learning baselines. Experimental results show that our method achieves superior performance with an AUC-ROC of 0.982, precision of 0.891, recall of 0.876, and F1-score of 0.883, outperforming the representative baselines considered in this study, including traditional machine learning models, standalone deep learning architectures, and supervised sequential neural models. Ablation studies confirm the individual contributions of both the sequential modeling component and the generative oversampling strategy. Our work demonstrates that combining temporal sequence modeling with generative synthesis provides a robust solution for imbalanced fraud detection, with potential applications extending to other domains requiring sequential pattern recognition under extreme class imbalance.

The proliferation of fake reviews on e-commerce and social platforms has severely undermined consumer trust and market integrity, necessitating robust and interpretable real-time detection mechanisms with multi-sensor data fusion capabilities. While traditional machine learning approaches have shown promise in identifying fraudulent reviews, they often lack transparency and fail to leverage the rich contextual knowledge embedded in large-scale datasets. In this paper, we propose KE-MLLM (Knowledge-Enhanced Multimodal Large Language Model), a unified framework that integrates knowledge-enhanced prompting with parameter-efficient fine-tuning for explainable fake review detection. Our approach employs LoRA (Low-Rank Adaptation) to fine-tune lightweight large language models (LLaMA-3-8B) on review text, while incorporating multimodal behavioral sensor signals including temporal patterns, user metadata, and social network characteristics for comprehensive anomaly sensing. To address the critical need for interpretability in fraud detection systems, we implement a Chain-of-Thought (CoT) reasoning module that generates human-understandable explanations for classification decisions, highlighting linguistic anomalies, sentiment inconsistencies, and behavioral red flags. We enhance the model’s discriminative capability through a knowledge distillation strategy that transfers domain-specific expertise from larger teacher models while maintaining computational efficiency suitable for edge sensing devices. Extensive experiments on two benchmark datasets—YelpChi and Amazon Reviews from the DGL Fraud Dataset—show that KE-MLLM achieves strong performance, reaching an F1-score of 94.3% and an AUC-ROC of 96.7% on YelpChi and outperforming the strongest baseline in our comparison by 5.8 and 4.2 percentage points, respectively. Furthermore, human evaluation indicates that the generated explanations achieve 89.5% consistency with expert annotations, suggesting that the framework can improve the interpretability and practical usefulness of automated fraud detection systems. The proposed framework provides a useful step toward more accurate and interpretable fake review detection and offers a practical reference for building more transparent and accountable AI systems in high-stakes applications.