CoAtNet-DeepMoE: A Convolution-Attention Hybrid with DeepSeek Mixture-of-Experts for Parameter-Efficient Tomato Disease Classification
为解决番茄疾病分类中模型参数过多的问题，提出了一种结合卷积-注意力机制与深度混合专家系统的CoAtNet-DeepMoE方法，在保持高精度的同时显著减少了参数量。
0 citationsRead paper
Selected work
Representative Papers
3D Gait-Based Autism Classification Using Attention-Enhanced Deep Learning with Cross-Fold Statistical Stability Analysis
本文针对自闭症早期诊断难题，采用注意力增强的深度学习方法分析3D步态数据，通过跨折统计稳定性分析提高分类准确性。
0 citationsRead paper
SAFER-Activities: A Dataset for Smart Assessment of Fall Events and Routine Activities
为了解决智能健康监控系统中动作识别精度不足的问题，特别是针对跌倒事件，通过创建包含详细帧级标注的SAFER-Activities数据集，并采用多种模型进行测试以提高识别准确性。
0 citationsRead paper
PolicyKG: An Agentic LLM Pipeline for Translating Institutional Policies into SHACL Knowledge Graphs
This work proposes a four-stage pipeline leveraging large language models to automatically translate institutionally authored natural language policies into machine-readable SHACL constraints for automated compliance checking. The approach integrates a LangGraph-based state machine, first-order deontic logic, and a YAML-based vocabulary registry, and introduces a novel plug-in Corpus Adapter mechanism that enables cross-domain transferability by simply swapping the vocabulary registry—eliminating the need for model retraining. Empirical findings reveal that higher-order logical constructs are exceedingly rare in institutional policies; on the AIT corpus, the method achieves 86.9% accuracy (κ = 0.709) in deontic classification and an F1 score of 0.866 for SHACL generation. Switching to a GDPR-specific registry significantly improves attribute alignment (p < 0.001), with fully reproducible results.
0 citationsRead paper
From Overload to Convergence: Supporting Multi-Issue Human-AI Negotiation with Bayesian Visualization
This study addresses the challenge of increased cognitive load in human–agent negotiation as the number of negotiation issues grows, which impairs both performance and autonomy. To mitigate this, the paper proposes the first decision support mechanism that integrates Bayesian estimation of agreement likelihood with interactive uncertainty visualization. Deployed in a residential lease negotiation scenario, the system dynamically visualizes the convergence of mutually acceptable agreement spaces, enabling users to efficiently identify high-potential options. Experimental results from 32 participants demonstrate that the approach significantly improves negotiation outcome quality and efficiency without redistributing bargaining surplus, while effectively preserving human negotiators’ sense of control. These findings underscore the method’s practical utility and novelty in supporting complex, multi-issue human–agent negotiations.
0 citationsRead paper
Recent publications
Latest Papers
CoAtNet-DeepMoE: A Convolution-Attention Hybrid with DeepSeek Mixture-of-Experts for Parameter-Efficient Tomato Disease Classification
为解决番茄疾病分类中模型参数过多的问题，提出了一种结合卷积-注意力机制与深度混合专家系统的CoAtNet-DeepMoE方法，在保持高精度的同时显著减少了参数量。
0 citationsRead paper
3D Gait-Based Autism Classification Using Attention-Enhanced Deep Learning with Cross-Fold Statistical Stability Analysis
本文针对自闭症早期诊断难题，采用注意力增强的深度学习方法分析3D步态数据，通过跨折统计稳定性分析提高分类准确性。
0 citationsRead paper
SAFER-Activities: A Dataset for Smart Assessment of Fall Events and Routine Activities
为了解决智能健康监控系统中动作识别精度不足的问题，特别是针对跌倒事件，通过创建包含详细帧级标注的SAFER-Activities数据集，并采用多种模型进行测试以提高识别准确性。
0 citationsRead paper
PolicyKG: An Agentic LLM Pipeline for Translating Institutional Policies into SHACL Knowledge Graphs
This work proposes a four-stage pipeline leveraging large language models to automatically translate institutionally authored natural language policies into machine-readable SHACL constraints for automated compliance checking. The approach integrates a LangGraph-based state machine, first-order deontic logic, and a YAML-based vocabulary registry, and introduces a novel plug-in Corpus Adapter mechanism that enables cross-domain transferability by simply swapping the vocabulary registry—eliminating the need for model retraining. Empirical findings reveal that higher-order logical constructs are exceedingly rare in institutional policies; on the AIT corpus, the method achieves 86.9% accuracy (κ = 0.709) in deontic classification and an F1 score of 0.866 for SHACL generation. Switching to a GDPR-specific registry significantly improves attribute alignment (p < 0.001), with fully reproducible results.
0 citationsRead paper
From Overload to Convergence: Supporting Multi-Issue Human-AI Negotiation with Bayesian Visualization
This study addresses the challenge of increased cognitive load in human–agent negotiation as the number of negotiation issues grows, which impairs both performance and autonomy. To mitigate this, the paper proposes the first decision support mechanism that integrates Bayesian estimation of agreement likelihood with interactive uncertainty visualization. Deployed in a residential lease negotiation scenario, the system dynamically visualizes the convergence of mutually acceptable agreement spaces, enabling users to efficiently identify high-potential options. Experimental results from 32 participants demonstrate that the approach significantly improves negotiation outcome quality and efficiency without redistributing bargaining surplus, while effectively preserving human negotiators’ sense of control. These findings underscore the method’s practical utility and novelty in supporting complex, multi-issue human–agent negotiations.
0 citationsRead paper