Selected work

Representative Papers

Context-Aware Autoencoders for Anomaly Detection in Maritime Surveillance

Jan 27, 2026

This work addresses the challenge of detecting collective and context-dependent anomalies in maritime surveillance, which traditional autoencoders struggle to capture due to their inability to leverage vessel-specific contextual information such as AIS messages. To overcome this limitation, the authors propose a context-aware autoencoder that, for the first time, integrates context-specific thresholds into the autoencoder framework, dynamically adjusting the reconstruction loss criterion. By jointly modeling contextual dependencies, temporal dynamics, and AIS data characteristics, the method significantly enhances the detection of anomalous fishing vessel behaviors. The approach not only reduces computational overhead but also outperforms conventional anomaly detection techniques, thereby demonstrating the critical role of contextual information in refining reconstruction error modeling and improving detection performance.

0 citationsRead paper

UWarp: A Whole Slide Image Registration Pipeline to Characterize Scanner-Induced Local Domain Shift

Mar 26, 2025

This work addresses patch-level domain shift induced by heterogeneous digital slide scanners in computational pathology. We propose the first local domain shift quantification framework specifically designed for computational pathology. Methodologically, we introduce a hierarchical registration architecture: global affine correction followed by local non-rigid deformation compensation, with registration accuracy quantitatively evaluated via target registration error (TRE). On two private datasets, median TRE is <4 pixels (<1 μm at 40× magnification), achieving significant efficiency gains. A key finding is a strong spatial correlation between model prediction variability and local tissue density. To our knowledge, this is the first study to enable interpretable, patch-level quantification of domain shift. Our framework establishes a novel paradigm and provides a practical tool for mitigating scanner-induced domain shift in digital pathology.

0 citationsRead paper

Risk-averse policies for natural gas futures trading using distributional reinforcement learning

Jan 08, 2025

Managing risk in highly volatile natural gas futures markets remains challenging. Method: This paper pioneers the application of distributional reinforcement learning (DRL)—specifically C51, QR-DQN, and IQN—to energy finance trading, integrating CVaR-optimized risk-sensitive policy learning with a tunable risk-preference mechanism. The approach jointly models return uncertainty via distributional RL, quantifies tail risk using CVaR, and leverages deep Q-networks for end-to-end decision-making. Results: Empirical evaluation shows C51 improves the return-risk ratio by over 32% compared to classical DQN; CVaR confidence level enables linear adjustment of risk aversion; and both C51 and IQN demonstrate superior robustness and adaptability across multi-regime volatility scenarios versus five state-of-the-art machine learning baselines. This work establishes an interpretable, controllable, risk-aware paradigm for intelligent trading in energy derivatives.

0 citationsRead paper

Assessing the impact of external factors on the occurrence of emergencies

Jan 06, 2025

This study investigates how weather, traffic, air quality, and temporal factors influence emergency medical service (EMS) call frequency, using dispatch data from Lausanne University Hospital, Switzerland, to support dynamic EMS resource allocation. Method: We integrate classical statistical tests—including correlation analysis, chi-square tests, t-tests, and information value—with interpretable AI techniques (SHAP and permutation importance), employing XGBoost and multilayer perceptron (MLP) models. Contribution/Results: We find that the “hour-of-day” feature alone achieves statistically significant predictive power (p < 0.01), and its single-feature model performs comparably to the full-feature model (p > 0.05). This challenges the necessity of complex multi-factor modeling for EMS demand forecasting. The result enables a lightweight, deployable, and robust real-time dispatch decision-support framework, offering a novel paradigm for intelligent EMS operations.

0 citationsRead paper

Recent publications

Latest Papers

Context-Aware Autoencoders for Anomaly Detection in Maritime Surveillance

Jan 27, 2026

This work addresses the challenge of detecting collective and context-dependent anomalies in maritime surveillance, which traditional autoencoders struggle to capture due to their inability to leverage vessel-specific contextual information such as AIS messages. To overcome this limitation, the authors propose a context-aware autoencoder that, for the first time, integrates context-specific thresholds into the autoencoder framework, dynamically adjusting the reconstruction loss criterion. By jointly modeling contextual dependencies, temporal dynamics, and AIS data characteristics, the method significantly enhances the detection of anomalous fishing vessel behaviors. The approach not only reduces computational overhead but also outperforms conventional anomaly detection techniques, thereby demonstrating the critical role of contextual information in refining reconstruction error modeling and improving detection performance.

0 citationsRead paper

UWarp: A Whole Slide Image Registration Pipeline to Characterize Scanner-Induced Local Domain Shift

Mar 26, 2025

This work addresses patch-level domain shift induced by heterogeneous digital slide scanners in computational pathology. We propose the first local domain shift quantification framework specifically designed for computational pathology. Methodologically, we introduce a hierarchical registration architecture: global affine correction followed by local non-rigid deformation compensation, with registration accuracy quantitatively evaluated via target registration error (TRE). On two private datasets, median TRE is <4 pixels (<1 μm at 40× magnification), achieving significant efficiency gains. A key finding is a strong spatial correlation between model prediction variability and local tissue density. To our knowledge, this is the first study to enable interpretable, patch-level quantification of domain shift. Our framework establishes a novel paradigm and provides a practical tool for mitigating scanner-induced domain shift in digital pathology.

0 citationsRead paper

Risk-averse policies for natural gas futures trading using distributional reinforcement learning

Jan 08, 2025

Managing risk in highly volatile natural gas futures markets remains challenging. Method: This paper pioneers the application of distributional reinforcement learning (DRL)—specifically C51, QR-DQN, and IQN—to energy finance trading, integrating CVaR-optimized risk-sensitive policy learning with a tunable risk-preference mechanism. The approach jointly models return uncertainty via distributional RL, quantifies tail risk using CVaR, and leverages deep Q-networks for end-to-end decision-making. Results: Empirical evaluation shows C51 improves the return-risk ratio by over 32% compared to classical DQN; CVaR confidence level enables linear adjustment of risk aversion; and both C51 and IQN demonstrate superior robustness and adaptability across multi-regime volatility scenarios versus five state-of-the-art machine learning baselines. This work establishes an interpretable, controllable, risk-aware paradigm for intelligent trading in energy derivatives.

0 citationsRead paper

Assessing the impact of external factors on the occurrence of emergencies

Jan 06, 2025

This study investigates how weather, traffic, air quality, and temporal factors influence emergency medical service (EMS) call frequency, using dispatch data from Lausanne University Hospital, Switzerland, to support dynamic EMS resource allocation. Method: We integrate classical statistical tests—including correlation analysis, chi-square tests, t-tests, and information value—with interpretable AI techniques (SHAP and permutation importance), employing XGBoost and multilayer perceptron (MLP) models. Contribution/Results: We find that the “hour-of-day” feature alone achieves statistically significant predictive power (p < 0.01), and its single-feature model performs comparably to the full-feature model (p > 0.05). This challenges the necessity of complex multi-factor modeling for EMS demand forecasting. The result enables a lightweight, deployable, and robust real-time dispatch decision-support framework, offering a novel paradigm for intelligent EMS operations.

0 citationsRead paper