Selected work

Representative Papers

LocaGen: Low-Overhead Indoor Localization Through Spatial Augmentation

Nov 22, 2025

Fingerprint-based indoor localization requires extensive on-site collection of labeled signal measurements, incurring high deployment costs. To address this, we propose a spatially enhanced conditional diffusion generative framework that synthesizes high-fidelity WiFi fingerprints for unobserved locations. Our method integrates density-driven location sampling, domain-informed data augmentation, and a spatially aware loss function. Crucially, it eliminates the need to collect measurements at all target locations, substantially reducing site-survey overhead. Evaluated on real-world WiFi datasets, our approach achieves localization accuracy comparable to the full-sampling baseline—even when only 70% of locations are observed—and outperforms existing generative methods by up to 28% in accuracy. The core contribution is the first incorporation of explicit spatial priors into conditional diffusion modeling, enabling physically interpretable and highly generalizable few-shot fingerprint synthesis.

0 citationsRead paper

Unsupervised Graph Neural Network Framework for Balanced Multipatterning in Advanced Electronic Design Automation Layouts

Nov 20, 2025

To address the dual challenges of conflict elimination and mask load balancing in multi-patterning decomposition for advanced technology nodes, this paper formulates the problem as a constrained graph coloring task and proposes an unsupervised graph neural network (GNN) framework. The method requires no labeled data and jointly optimizes the primary objective—achieving zero conflicts—and the secondary objective—balancing feature counts across masks. It first employs a GNN to generate high-quality initial colorings, then refines them iteratively via a GNN-guided heuristic strategy integrated with simulated annealing. Evaluated on both private and public layout datasets, the approach achieves 100% zero-conflict decompositions while significantly improving color distribution balance—reducing standard deviation by ≥40%. The framework is reproducible, computationally efficient, and demonstrates strong potential for industrial deployment.

0 citationsRead paper

CephRes-MHNet: A Multi-Head Residual Network for Accurate and Robust Cephalometric Landmark Detection

Nov 13, 2025

This study addresses the challenge of automatic craniofacial landmark localization in 2D lateral cephalometric X-rays—where manual annotation is time-consuming and error-prone, and existing methods lack robustness under low contrast and anatomical complexity. We propose a lightweight Multi-Head Residual Network (MHR-Net), integrating residual encoding, dual channel-spatial attention mechanisms, and a multi-head decoding architecture to significantly enhance anatomical context modeling and keypoint localization accuracy. Trained end-to-end on the Aariz dataset (1,000 cases), MHR-Net achieves a state-of-the-art mean radial error of 1.23 mm and an 85.5% success rate within 2.0 mm, using less than 25% of the parameters of the strongest baseline. The method establishes a new paradigm for clinical cephalometric analysis, offering both high accuracy and computational efficiency.

0 citationsRead paper

Federated Learning Meets LLMs: Feature Extraction From Heterogeneous Clients

Sep 29, 2025

To address the challenge of aligning multi-source tabular data in federated learning—where schema heterogeneity and incompatible feature spaces impede effective integration—this paper proposes a large language model (LLM)-based semantic alignment framework. Clients locally serialize tabular data into textual sequences and extract privacy-preserving semantic embeddings using lightweight pre-trained LLMs (e.g., DistilBERT, ALBERT, RoBERTa, ClinicalBERT), enabling fully automated, human-in-the-loop-free feature alignment. A lightweight classifier is then trained via FedAvg. Unlike conventional schema-matching approaches, this method eliminates explicit schema coordination, substantially enhancing robustness and generalization under heterogeneity. Evaluated on coronary heart disease prediction, it achieves up to a 0.25 improvement in F1-score, reduces communication overhead by 65%, and maintains stable performance even under extreme schema divergence.

0 citationsRead paper

ViG-LRGC: Vision Graph Neural Networks with Learnable Reparameterized Graph Construction

Sep 23, 2025

This work addresses the reliance of Vision Graph Neural Networks (ViGs) on hand-crafted hyperparameters or fixed rules for graph construction. To overcome this limitation, we propose Learnable Reparameterized Graph Construction (LRGC), which models images as dynamic node graphs. At each layer, node relationships are computed via key-query attention, and soft-threshold-based reparameterization enables end-to-end differentiable edge selection—automatically learning optimal connection thresholds while avoiding biases introduced by clustering or hard thresholding. LRGC requires no hyperparameter tuning, achieving both structural adaptivity and full differentiability during training. Evaluated on ImageNet-1K, LRGC significantly outperforms state-of-the-art ViG models of comparable capacity, demonstrating superior effectiveness and generalization in image representation learning and classification tasks.

0 citationsRead paper

Recent publications

Latest Papers

LocaGen: Low-Overhead Indoor Localization Through Spatial Augmentation

Nov 22, 2025

Fingerprint-based indoor localization requires extensive on-site collection of labeled signal measurements, incurring high deployment costs. To address this, we propose a spatially enhanced conditional diffusion generative framework that synthesizes high-fidelity WiFi fingerprints for unobserved locations. Our method integrates density-driven location sampling, domain-informed data augmentation, and a spatially aware loss function. Crucially, it eliminates the need to collect measurements at all target locations, substantially reducing site-survey overhead. Evaluated on real-world WiFi datasets, our approach achieves localization accuracy comparable to the full-sampling baseline—even when only 70% of locations are observed—and outperforms existing generative methods by up to 28% in accuracy. The core contribution is the first incorporation of explicit spatial priors into conditional diffusion modeling, enabling physically interpretable and highly generalizable few-shot fingerprint synthesis.

0 citationsRead paper

Unsupervised Graph Neural Network Framework for Balanced Multipatterning in Advanced Electronic Design Automation Layouts

Nov 20, 2025

To address the dual challenges of conflict elimination and mask load balancing in multi-patterning decomposition for advanced technology nodes, this paper formulates the problem as a constrained graph coloring task and proposes an unsupervised graph neural network (GNN) framework. The method requires no labeled data and jointly optimizes the primary objective—achieving zero conflicts—and the secondary objective—balancing feature counts across masks. It first employs a GNN to generate high-quality initial colorings, then refines them iteratively via a GNN-guided heuristic strategy integrated with simulated annealing. Evaluated on both private and public layout datasets, the approach achieves 100% zero-conflict decompositions while significantly improving color distribution balance—reducing standard deviation by ≥40%. The framework is reproducible, computationally efficient, and demonstrates strong potential for industrial deployment.

0 citationsRead paper

CephRes-MHNet: A Multi-Head Residual Network for Accurate and Robust Cephalometric Landmark Detection

Nov 13, 2025

This study addresses the challenge of automatic craniofacial landmark localization in 2D lateral cephalometric X-rays—where manual annotation is time-consuming and error-prone, and existing methods lack robustness under low contrast and anatomical complexity. We propose a lightweight Multi-Head Residual Network (MHR-Net), integrating residual encoding, dual channel-spatial attention mechanisms, and a multi-head decoding architecture to significantly enhance anatomical context modeling and keypoint localization accuracy. Trained end-to-end on the Aariz dataset (1,000 cases), MHR-Net achieves a state-of-the-art mean radial error of 1.23 mm and an 85.5% success rate within 2.0 mm, using less than 25% of the parameters of the strongest baseline. The method establishes a new paradigm for clinical cephalometric analysis, offering both high accuracy and computational efficiency.

0 citationsRead paper

Federated Learning Meets LLMs: Feature Extraction From Heterogeneous Clients

Sep 29, 2025

To address the challenge of aligning multi-source tabular data in federated learning—where schema heterogeneity and incompatible feature spaces impede effective integration—this paper proposes a large language model (LLM)-based semantic alignment framework. Clients locally serialize tabular data into textual sequences and extract privacy-preserving semantic embeddings using lightweight pre-trained LLMs (e.g., DistilBERT, ALBERT, RoBERTa, ClinicalBERT), enabling fully automated, human-in-the-loop-free feature alignment. A lightweight classifier is then trained via FedAvg. Unlike conventional schema-matching approaches, this method eliminates explicit schema coordination, substantially enhancing robustness and generalization under heterogeneity. Evaluated on coronary heart disease prediction, it achieves up to a 0.25 improvement in F1-score, reduces communication overhead by 65%, and maintains stable performance even under extreme schema divergence.

0 citationsRead paper

ViG-LRGC: Vision Graph Neural Networks with Learnable Reparameterized Graph Construction

Sep 23, 2025

This work addresses the reliance of Vision Graph Neural Networks (ViGs) on hand-crafted hyperparameters or fixed rules for graph construction. To overcome this limitation, we propose Learnable Reparameterized Graph Construction (LRGC), which models images as dynamic node graphs. At each layer, node relationships are computed via key-query attention, and soft-threshold-based reparameterization enables end-to-end differentiable edge selection—automatically learning optimal connection thresholds while avoiding biases introduced by clustering or hard thresholding. LRGC requires no hyperparameter tuning, achieving both structural adaptivity and full differentiability during training. Evaluated on ImageNet-1K, LRGC significantly outperforms state-of-the-art ViG models of comparable capacity, demonstrating superior effectiveness and generalization in image representation learning and classification tasks.

0 citationsRead paper