Transfer Learning on Edge Connecting Probability Estimation Under Graphon Model NeurIPS 2025 · Code
Yuyao Wang, Yu-Hung Cheng, Debarghya Mukherjee, Huimin Cheng
Boston University

• We propose GTRANS, the first graphon transfer learning method without node correspondence, combining Gromov-Wasserstein optimal transport and residual smoothing.
• Theoretical Guarantee: We prove stability and convergence of the transport-based alignment under nonparametric assumptions.
• Applications: GTRANS improves link prediction and graph classification, especially under small target graphs and sparse settings.
• Achieves state-of-the-art (SOTA) performance on both synthetic and real-world datasets.

Phase Transition in Nonparametric Minimax Rates for Covariate Shifts on Approximate Manifolds · Code · Poster · Presentation
Yuyao Wang, Nabarun Deb, Debarghya Mukherjee
Boston University; The University of Chicago Booth School of Business

• We establish new minimax rates for estimating Hölder-smooth regression functions under covariate shift when the target distribution lies near, but not on, a source manifold.
• Introduces a novel phase transition phenomenon: the minimax rate depends sharply on the proximity between the target and source support, unifying prior results under a generalized Hölder framework.
• Addresses settings where density ratios are ill-defined, making classical transfer techniques invalid.
• Our estimator adapts to unknown manifold dimension and achieves near-optimal rates without prior geometric knowledge.

Cross-Domain Hyperspectral Image Classification via Mamba-CNN and Knowledge Distillation · Presentation
Aoyan Du, Guixin Zhao, Yuyao Wang, Aimei Dong, Guohua Lv, Yongbiao Gao, Xiangjun Dong
Shandong Computer Science Center; Boston University

• We propose MKDnet, a cross-domain HSI classification framework combining Mamba-CNN hybrid architecture with knowledge distillation and graph alignment via optimal transport.
• Effectively captures both global context and local detail via Mamba + CNN dual-stream encoder.
• Aligns source–target distributions through domain-level knowledge distillation and graph OT-based subgraph matching.
• Achieves SOTA performance across multiple public hyperspectral benchmarks under domain shift.

Multi-scale based Cross-modal Semantic Alignment Network for Radiology Report Generation
Zhihao Zhang, Long Zhao, Yuyao Wang, Dun Lan, Linfeng Jiang, Xiangjun Dong
Shandong Computer Science Center; Boston University

• We propose MCSANet, a radiology report generation framework that enhances cross-modal semantic alignment between medical images and diagnostic text.
• Introduces a Multi-scale Visual Feature Extraction (MVE) module with multi-head local sparse attention (MLSA) to capture image semantics and abnormalities across different spatial scales.
• Incorporates a Cross-modal Semantic Alignment (CSA) module with a learnable matrix, gating mechanism, and multi-label contrastive loss for precise image–text fusion.
• Combined with a Transformer-based report generator, MCSANet achieves SOTA performance on IU-Xray and MIMIC-CXR, surpassing prior models such as CAMANet and XPRONET.