Selected work

Representative Papers

Federated Inference: Toward Privacy-Preserving Collaborative and Incentivized Model Serving

Feb 09, 2026

This work proposes Federated Inference (FI) as a complementary paradigm to federated learning, enabling secure collaboration among private models during inference without sharing data or model parameters. The study introduces the first unified abstraction framework for FI, articulating two core objectives: preserving privacy during inference and achieving performance gains through collaboration. Building upon secure multi-party computation, the authors design a privacy-preserving collaborative inference architecture that integrates ensemble learning and incentive mechanisms. Systematic modeling and empirical analysis are conducted under non-IID data distributions and stringent privacy constraints. Experimental results reveal critical trade-offs among privacy, collaboration efficacy, and incentive alignment, underscoring the necessity of designing FI systems independently from conventional training-centric paradigms.

0 citationsRead paper

Context-aware Graph Causality Inference for Few-Shot Molecular Property Prediction

Jan 16, 2026

This work addresses the challenge of leveraging functional group–based causal priors and identifying critical substructures in few-shot molecular property prediction. To this end, we propose CaMol, a novel framework that introduces causal inference into this task for the first time. By constructing a contextual graph that integrates functional groups, molecular structures, and target properties, and combining it with a learnable atomic masking strategy and a chemistry-informed backdoor adjustment mechanism, CaMol effectively disentangles causal effects from confounding factors to identify substructures directly causally linked to the target property. Extensive experiments demonstrate that CaMol significantly improves both prediction accuracy and sample efficiency across multiple datasets. Moreover, the identified causal substructures show strong alignment with known functional groups, highlighting the model’s high performance and interpretability.

0 citationsRead paper

Pre-training Graph Neural Networks on 2D and 3D Molecular Structures by using Multi-View Conditional Information Bottleneck

Nov 23, 2025

Existing molecular graph pretraining struggles to simultaneously ensure cross-view (2D/3D) semantic consistency and alignment of functionally critical substructures. To address this, we propose the Multi-View Conditional Information Bottleneck (MVCIB) framework, which achieves fine-grained cross-view alignment via context-guided representation learning and a functional-group-driven substructure anchoring mechanism. MVCIB is the first method to attain geometric discriminability at the 3D Weisfeiler–Lehman hierarchy level. It integrates ego-network modeling with cross-view attention to jointly optimize 2D topological and 3D geometric representations. Evaluated on four molecular property prediction benchmarks, MVCIB significantly outperforms state-of-the-art methods—particularly excelling at distinguishing stereoisomers with identical 2D graphs but distinct 3D conformations. The framework delivers both superior predictive performance and enhanced interpretability through principled substructure grounding.

0 citationsRead paper

Image Enhancement Based on Pigment Representation

Oct 03, 2025

Existing image enhancement methods are constrained by fixed, predefined color spaces (e.g., RGB), lacking content adaptivity and thus struggling to balance enhancement quality and computational efficiency. To address this, we propose a novel image enhancement framework based on learnable, high-dimensional pigment representations. Our method first employs a vision encoder to predict content-adaptive pigment transformation and reprojection parameters; then decomposes the RGB input into high-dimensional pigment features, performs pigment-wise reprojection and fusion; and finally reconstructs the enhanced result via inverse transformation to RGB space. This work pioneers the modeling of pigment representation as a learnable, high-dimensional, content-driven feature space—breaking free from conventional color space limitations. Extensive experiments demonstrate state-of-the-art performance on image retouching and tone mapping tasks, while maintaining low computational overhead and a lightweight architecture (<1.2M parameters).

0 citationsRead paper

Discovering Spatial Correlations between Earth Observations in Global Atmospheric State Estimation by using Adaptive Graph Structure Learning

Aug 11, 2025

Existing global atmospheric state estimation methods inadequately model spatial correlations between Earth observations and atmospheric fields. Method: We propose an adaptive spatiotemporal graph neural network (STGNN) framework that introduces an edge-sampling mechanism jointly guided by node-degree adaptivity and spatial-distance constraints to mitigate information loss and over-smoothing in graph learning; further, it jointly models meteorological observations and numerical weather prediction (NWP) gridded data to explicitly capture dynamic spatiotemporal dependencies. Contribution/Results: Evaluated on real-world observational data across East Asia, our model significantly outperforms state-of-the-art STGNN approaches—particularly in regions with sharp atmospheric transitions (e.g., frontal zones and typhoons), where forecast accuracy improves markedly. The framework establishes a new, interpretable, and robust paradigm for high-resolution atmospheric state estimation.

0 citationsRead paper

Recent publications

Latest Papers

Federated Inference: Toward Privacy-Preserving Collaborative and Incentivized Model Serving

Feb 09, 2026

This work proposes Federated Inference (FI) as a complementary paradigm to federated learning, enabling secure collaboration among private models during inference without sharing data or model parameters. The study introduces the first unified abstraction framework for FI, articulating two core objectives: preserving privacy during inference and achieving performance gains through collaboration. Building upon secure multi-party computation, the authors design a privacy-preserving collaborative inference architecture that integrates ensemble learning and incentive mechanisms. Systematic modeling and empirical analysis are conducted under non-IID data distributions and stringent privacy constraints. Experimental results reveal critical trade-offs among privacy, collaboration efficacy, and incentive alignment, underscoring the necessity of designing FI systems independently from conventional training-centric paradigms.

0 citationsRead paper

Context-aware Graph Causality Inference for Few-Shot Molecular Property Prediction

Jan 16, 2026

This work addresses the challenge of leveraging functional group–based causal priors and identifying critical substructures in few-shot molecular property prediction. To this end, we propose CaMol, a novel framework that introduces causal inference into this task for the first time. By constructing a contextual graph that integrates functional groups, molecular structures, and target properties, and combining it with a learnable atomic masking strategy and a chemistry-informed backdoor adjustment mechanism, CaMol effectively disentangles causal effects from confounding factors to identify substructures directly causally linked to the target property. Extensive experiments demonstrate that CaMol significantly improves both prediction accuracy and sample efficiency across multiple datasets. Moreover, the identified causal substructures show strong alignment with known functional groups, highlighting the model’s high performance and interpretability.

0 citationsRead paper

Pre-training Graph Neural Networks on 2D and 3D Molecular Structures by using Multi-View Conditional Information Bottleneck

Nov 23, 2025

Existing molecular graph pretraining struggles to simultaneously ensure cross-view (2D/3D) semantic consistency and alignment of functionally critical substructures. To address this, we propose the Multi-View Conditional Information Bottleneck (MVCIB) framework, which achieves fine-grained cross-view alignment via context-guided representation learning and a functional-group-driven substructure anchoring mechanism. MVCIB is the first method to attain geometric discriminability at the 3D Weisfeiler–Lehman hierarchy level. It integrates ego-network modeling with cross-view attention to jointly optimize 2D topological and 3D geometric representations. Evaluated on four molecular property prediction benchmarks, MVCIB significantly outperforms state-of-the-art methods—particularly excelling at distinguishing stereoisomers with identical 2D graphs but distinct 3D conformations. The framework delivers both superior predictive performance and enhanced interpretability through principled substructure grounding.

0 citationsRead paper

Image Enhancement Based on Pigment Representation

Oct 03, 2025

Existing image enhancement methods are constrained by fixed, predefined color spaces (e.g., RGB), lacking content adaptivity and thus struggling to balance enhancement quality and computational efficiency. To address this, we propose a novel image enhancement framework based on learnable, high-dimensional pigment representations. Our method first employs a vision encoder to predict content-adaptive pigment transformation and reprojection parameters; then decomposes the RGB input into high-dimensional pigment features, performs pigment-wise reprojection and fusion; and finally reconstructs the enhanced result via inverse transformation to RGB space. This work pioneers the modeling of pigment representation as a learnable, high-dimensional, content-driven feature space—breaking free from conventional color space limitations. Extensive experiments demonstrate state-of-the-art performance on image retouching and tone mapping tasks, while maintaining low computational overhead and a lightweight architecture (<1.2M parameters).

0 citationsRead paper

Discovering Spatial Correlations between Earth Observations in Global Atmospheric State Estimation by using Adaptive Graph Structure Learning

Aug 11, 2025

Existing global atmospheric state estimation methods inadequately model spatial correlations between Earth observations and atmospheric fields. Method: We propose an adaptive spatiotemporal graph neural network (STGNN) framework that introduces an edge-sampling mechanism jointly guided by node-degree adaptivity and spatial-distance constraints to mitigate information loss and over-smoothing in graph learning; further, it jointly models meteorological observations and numerical weather prediction (NWP) gridded data to explicitly capture dynamic spatiotemporal dependencies. Contribution/Results: Evaluated on real-world observational data across East Asia, our model significantly outperforms state-of-the-art STGNN approaches—particularly in regions with sharp atmospheric transitions (e.g., frontal zones and typhoons), where forecast accuracy improves markedly. The framework establishes a new, interpretable, and robust paradigm for high-resolution atmospheric state estimation.

0 citationsRead paper