Selected work

Representative Papers

Weak Relation Enforcement for Kinematic-Informed Long-Term Stock Prediction with Artificial Neural Networks

Nov 13, 2025Sai

Artificial neural networks (ANNs) suffer from spurious predictions in long-horizon stock forecasting due to high data volatility, out-of-distribution (OOD) test samples, and outliers. Method: This paper proposes a kinematics-inspired weak-relation-constrained neural network. It jointly optimizes both predictions and their first-order differences (“velocity”) in the loss function to explicitly model dynamic temporal evolution. A weak-relation enforcement mechanism mitigates structural distortion induced by autoregressive normalization while preserving original neighborhood topology. Additionally, a velocity-aware composite loss and normalization-sensitive activation functions are introduced, ensuring compatibility with diverse RNN architectures. Contribution/Results: Evaluated on 15 years of Dow Jones Industrial Average data, the method significantly improves long-term forecasting stability and statistical significance. It demonstrates superior robustness—particularly under OOD conditions and during periods of high market volatility—while maintaining architectural flexibility and interpretability through physics-informed constraints.

1 citationsRead paper

Elements of Active Continuous Learning and Uncertainty Self-awareness: A Narrow Implementation for Face and Facial Expression Recognition

Nov 04, 2025Artificial General Intelligence

This study addresses the lack of uncertainty awareness in foundational neural networks (e.g., CNNs for face and expression recognition). We propose a bi-level architecture that emulates “self-awareness”: a supervised artificial neural network (ANN) serves as a meta-monitor, analyzing the lower-layer CNN’s activation patterns in real time to dynamically estimate prediction confidence; upon detecting high uncertainty, it autonomously triggers active learning to request human annotation. Crucially, this is the first work to employ a trainable, memory-augmented supervised ANN to model the cognitive state of the model itself, thereby closing an uncertainty-driven adaptive decision loop. Experiments demonstrate substantial improvements in robustness and accuracy for face recognition and facial expression analysis under complex, real-world conditions. The approach advances AI trustworthiness and human-AI collaborative intelligence by enabling self-regulated, uncertainty-informed inference.

1 citationsRead paper

K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations

Sep 14, 2026

为评估大型语言模型在高风险心理健康对话中的安全性，开发了K-Bench基准测试，通过临床校准和多情景模拟来评价模型表现。

0 citationsRead paper

Sample Complexity and Decision-Theoretic Guarantees for Bayesian Model Averaging over Decision Trees with Catalan-Exponential Priors

May 31, 2026

This study investigates under what conditions Bayesian Model Averaging (BMA) in Bayesian decision trees yields sufficient epistemic confidence to justify deterministic decisions. Focusing on Bayesian decision trees equipped with Dirichlet–Multinomial leaf models and a Catalan-exponential prior over tree size, the work establishes—for the first time—theoretical non-asymptotic rational commitment thresholds in closed form. These thresholds precisely characterize the quantitative relationship between sample complexity and decision reliability in finite-sample settings. The analysis provides the first non-asymptotic decision-theoretic guarantee for the safe and effective use of BMA in making high-stakes decisions under limited data, thereby bridging a critical gap between Bayesian model averaging theory and practical decision-making requirements.

0 citationsRead paper

Reproducibility in Event-Log Research: A Parametrised Generator and Benchmark for Event-based Signatures

Jan 19, 2026

This study addresses the challenge of evaluating signature-based cybersecurity detection methods due to the scarcity of publicly available, real-world event logs, which are often restricted for privacy and sensitivity reasons. To overcome this limitation, the authors propose a parameterized synthetic event log generation approach that models known attack signatures to produce labeled, configurable log data reflecting realistic scenarios. Complementing this generator, they introduce a benchmarking framework specifically designed for signature-based detection evaluation. This framework establishes, for the first time, a reproducible and comparable testing environment. Experimental results demonstrate that on the generated benchmark datasets, the DBSCAN clustering algorithm achieves an Adjusted Rand Index exceeding 0.95 in most scenarios, thereby validating both the effectiveness and practical utility of the proposed methodology.

0 citationsRead paper

Recent publications

Latest Papers

K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations

Sep 14, 2026

为评估大型语言模型在高风险心理健康对话中的安全性，开发了K-Bench基准测试，通过临床校准和多情景模拟来评价模型表现。

0 citationsRead paper

Sample Complexity and Decision-Theoretic Guarantees for Bayesian Model Averaging over Decision Trees with Catalan-Exponential Priors

May 31, 2026

This study investigates under what conditions Bayesian Model Averaging (BMA) in Bayesian decision trees yields sufficient epistemic confidence to justify deterministic decisions. Focusing on Bayesian decision trees equipped with Dirichlet–Multinomial leaf models and a Catalan-exponential prior over tree size, the work establishes—for the first time—theoretical non-asymptotic rational commitment thresholds in closed form. These thresholds precisely characterize the quantitative relationship between sample complexity and decision reliability in finite-sample settings. The analysis provides the first non-asymptotic decision-theoretic guarantee for the safe and effective use of BMA in making high-stakes decisions under limited data, thereby bridging a critical gap between Bayesian model averaging theory and practical decision-making requirements.

0 citationsRead paper

Reproducibility in Event-Log Research: A Parametrised Generator and Benchmark for Event-based Signatures

Jan 19, 2026

This study addresses the challenge of evaluating signature-based cybersecurity detection methods due to the scarcity of publicly available, real-world event logs, which are often restricted for privacy and sensitivity reasons. To overcome this limitation, the authors propose a parameterized synthetic event log generation approach that models known attack signatures to produce labeled, configurable log data reflecting realistic scenarios. Complementing this generator, they introduce a benchmarking framework specifically designed for signature-based detection evaluation. This framework establishes, for the first time, a reproducible and comparable testing environment. Experimental results demonstrate that on the generated benchmark datasets, the DBSCAN clustering algorithm achieves an Adjusted Rand Index exceeding 0.95 in most scenarios, thereby validating both the effectiveness and practical utility of the proposed methodology.

0 citationsRead paper

Weak Relation Enforcement for Kinematic-Informed Long-Term Stock Prediction with Artificial Neural Networks

Nov 13, 2025Sai

Artificial neural networks (ANNs) suffer from spurious predictions in long-horizon stock forecasting due to high data volatility, out-of-distribution (OOD) test samples, and outliers. Method: This paper proposes a kinematics-inspired weak-relation-constrained neural network. It jointly optimizes both predictions and their first-order differences (“velocity”) in the loss function to explicitly model dynamic temporal evolution. A weak-relation enforcement mechanism mitigates structural distortion induced by autoregressive normalization while preserving original neighborhood topology. Additionally, a velocity-aware composite loss and normalization-sensitive activation functions are introduced, ensuring compatibility with diverse RNN architectures. Contribution/Results: Evaluated on 15 years of Dow Jones Industrial Average data, the method significantly improves long-term forecasting stability and statistical significance. It demonstrates superior robustness—particularly under OOD conditions and during periods of high market volatility—while maintaining architectural flexibility and interpretability through physics-informed constraints.

1 citationsRead paper

Batch Transformer Architecture: Case of Synthetic Image Generation for Emotion Expression Facial Recognition

Nov 13, 2025

To address poor generalization in facial expression recognition caused by scarce labeled data under challenging conditions such as makeup and occlusion, this paper proposes an Implicit Sparse-Style Batch Transformer architecture. Our method integrates a dimension-aware sparse attention mechanism within an encoder-decoder framework, focusing exclusively on principal-feature dimensions to drastically reduce bottleneck-layer parameters. By jointly optimizing feature selection and implicit style modeling, it enhances both semantic consistency and diversity of synthesized images. Evaluation on small-scale datasets of makeup- and occlusion-affected faces demonstrates that the generated samples effectively augment training data variability, boosting downstream expression recognition accuracy by 4.2% and significantly improving model robustness. The core contribution is the first introduction of an implicit sparse attention paradigm tailored to low-dimensional critical features—achieving a favorable trade-off among computational efficiency, interpretability, and generative quality.

0 citationsRead paper