Selected work

Representative Papers

Gromov-Wasserstein Methods for Multi-View Relational Embedding and Clustering

Apr 26, 2026

This work addresses the challenge of learning a unified low-dimensional representation from multi-view relational data, where inconsistent underlying geometric structures across views hinder effective integration. To overcome this, the authors propose a consensus embedding framework based on the Gromov–Wasserstein (GW) distance, which operates directly on pairwise distance matrices to preserve relational structures shared across views. By fusing intrinsic distances from multiple views and incorporating a clustering-oriented low-support representation, they introduce two novel algorithms—Bary-GWMDS and Mean-GWMDS-C—that robustly handle nonlinear distortions and yield geometrically consistent embeddings. Experimental results on both synthetic and real-world datasets demonstrate that the proposed methods produce representations with clear geometric interpretability and achieve superior clustering performance.

0 citationsRead paper

Structure-Preserving Multi-View Embedding Using Gromov-Wasserstein Optimal Transport

Apr 02, 2026

This work addresses the challenge of effectively fusing multi-view data under heterogeneous geometries or nonlinear distortions, where traditional methods often fail to recover a consistent low-dimensional structure. The study introduces Gromov–Wasserstein (GW) optimal transport into multi-view embedding for the first time, proposing two geometry-aware strategies. Mean-GWMDS aligns distance matrices from individual views via GW coupling, averages them, and applies multidimensional scaling to obtain a unified embedding. Multi-GWMDS, in contrast, generates multiple geometrically consistent candidate embeddings and selects the optimal one. Notably, both approaches operate without explicit feature alignment or concatenation, enabling robust handling of nonlinearities and heterogeneous geometries. Experiments on synthetic manifolds and real-world datasets demonstrate superior cross-view structural preservation compared to existing methods.

0 citationsRead paper

ReBaPL: Repulsive Bayesian Prompt Learning

Nov 21, 2025

Traditional prompt learning suffers from overfitting and poor out-of-distribution generalization. To address this, we propose Repulsive Bayesian Prompt Learning (ReBaPL), a Bayesian inference framework that models the posterior distribution over prompt parameters. ReBaPL introduces repulsive potential functions—based on Maximum Mean Discrepancy (MMD) and Wasserstein distance—to explicitly enforce diversity in the representation space. It further employs Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) with cyclical step-size scheduling to enable efficient multi-modal posterior sampling, effectively mitigating mode collapse. ReBaPL is modular and compatible with existing prompt-learning methods in a plug-and-play manner. Empirical evaluation across multiple benchmarks demonstrates substantial improvements in both generalization performance and posterior representation quality. By unifying uncertainty-aware prompting with structured posterior regularization, ReBaPL establishes a new paradigm for robust and scalable prompt learning.

0 citationsRead paper

Computing Wasserstein Barycenters through Gradient Flows

Oct 06, 2025

Existing discrete Wasserstein barycenter algorithms suffer from poor scalability and require full-sample access. This paper proposes a novel modeling framework based on the Wasserstein gradient flow, reformulating barycenter computation as an energy minimization problem in the space of probability measures—naturally incorporating geometric structure and enabling explicit energy-based regularization. The method employs a minibatch sampling scheme, drastically reducing computational and memory overhead. Convergence is theoretically guaranteed via analysis leveraging the Polyak–Łojasiewicz inequality. Experiments on synthetic datasets and domain adaptation tasks demonstrate superior accuracy, efficiency, and robustness compared to state-of-the-art discrete methods and neural-network baselines. Key contributions include: (i) the first scalable gradient-flow paradigm for Wasserstein barycenters; (ii) interpretable, energy-driven regularization; and (iii) establishing a new standard for minibatch-based Wasserstein barycenter computation.

0 citationsRead paper

Recent publications

Latest Papers

Gromov-Wasserstein Methods for Multi-View Relational Embedding and Clustering

Apr 26, 2026

This work addresses the challenge of learning a unified low-dimensional representation from multi-view relational data, where inconsistent underlying geometric structures across views hinder effective integration. To overcome this, the authors propose a consensus embedding framework based on the Gromov–Wasserstein (GW) distance, which operates directly on pairwise distance matrices to preserve relational structures shared across views. By fusing intrinsic distances from multiple views and incorporating a clustering-oriented low-support representation, they introduce two novel algorithms—Bary-GWMDS and Mean-GWMDS-C—that robustly handle nonlinear distortions and yield geometrically consistent embeddings. Experimental results on both synthetic and real-world datasets demonstrate that the proposed methods produce representations with clear geometric interpretability and achieve superior clustering performance.

0 citationsRead paper

Structure-Preserving Multi-View Embedding Using Gromov-Wasserstein Optimal Transport

Apr 02, 2026

This work addresses the challenge of effectively fusing multi-view data under heterogeneous geometries or nonlinear distortions, where traditional methods often fail to recover a consistent low-dimensional structure. The study introduces Gromov–Wasserstein (GW) optimal transport into multi-view embedding for the first time, proposing two geometry-aware strategies. Mean-GWMDS aligns distance matrices from individual views via GW coupling, averages them, and applies multidimensional scaling to obtain a unified embedding. Multi-GWMDS, in contrast, generates multiple geometrically consistent candidate embeddings and selects the optimal one. Notably, both approaches operate without explicit feature alignment or concatenation, enabling robust handling of nonlinearities and heterogeneous geometries. Experiments on synthetic manifolds and real-world datasets demonstrate superior cross-view structural preservation compared to existing methods.

0 citationsRead paper

ReBaPL: Repulsive Bayesian Prompt Learning

Nov 21, 2025

Traditional prompt learning suffers from overfitting and poor out-of-distribution generalization. To address this, we propose Repulsive Bayesian Prompt Learning (ReBaPL), a Bayesian inference framework that models the posterior distribution over prompt parameters. ReBaPL introduces repulsive potential functions—based on Maximum Mean Discrepancy (MMD) and Wasserstein distance—to explicitly enforce diversity in the representation space. It further employs Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) with cyclical step-size scheduling to enable efficient multi-modal posterior sampling, effectively mitigating mode collapse. ReBaPL is modular and compatible with existing prompt-learning methods in a plug-and-play manner. Empirical evaluation across multiple benchmarks demonstrates substantial improvements in both generalization performance and posterior representation quality. By unifying uncertainty-aware prompting with structured posterior regularization, ReBaPL establishes a new paradigm for robust and scalable prompt learning.

0 citationsRead paper

Computing Wasserstein Barycenters through Gradient Flows

Oct 06, 2025

Existing discrete Wasserstein barycenter algorithms suffer from poor scalability and require full-sample access. This paper proposes a novel modeling framework based on the Wasserstein gradient flow, reformulating barycenter computation as an energy minimization problem in the space of probability measures—naturally incorporating geometric structure and enabling explicit energy-based regularization. The method employs a minibatch sampling scheme, drastically reducing computational and memory overhead. Convergence is theoretically guaranteed via analysis leveraging the Polyak–Łojasiewicz inequality. Experiments on synthetic datasets and domain adaptation tasks demonstrate superior accuracy, efficiency, and robustness compared to state-of-the-art discrete methods and neural-network baselines. Key contributions include: (i) the first scalable gradient-flow paradigm for Wasserstein barycenters; (ii) interpretable, energy-driven regularization; and (iii) establishing a new standard for minibatch-based Wasserstein barycenter computation.

0 citationsRead paper