Transfer Learning on Edge Connecting Probability Estimation Under Graphon Model · NeurIPS · Poster · Slides · Code
Boston University

• Graphon transfer without node correspondence: aligns source/target graphs via Gromov–Wasserstein, then transfers edge-probability structure in a fully nonparametric way.
• Provably stable + practically strong: residual smoothing boosts small/sparse/heterogeneous targets, with convergence & stability guarantees and SOTA gains on link prediction / graph classification.

SCOT: Multi-Source Cross-City Transfer with Optimal-Transport Soft-Correspondence Objectives
Boston University; Shandong University

• Explicit many-to-many correspondence (no node matching): learns a Sinkhorn entropic-OT coupling between unequal region sets and uses it to define soft aligned pairs for cross-city transfer.
• OT-guided semantic sharpening + stability: combines OT-weighted contrastive alignment with cycle-style reconstruction; extends to multi-source with a target-aware prototype hub to prevent collapse and handle strong heterogeneity.

From Text to Forecasts: Bridging Modality Gap with Temporal Evolution Semantic Space
Boston University; Hong Kong University of Science and Technology; Shandong University

• Bridges the text–time-series modality gap: introduces a Temporal Evolution Semantic Space (TESS) that distills free-form text into interpretable temporal primitives (mean shift, volatility, shape, lag), instead of directly fusing noisy token embeddings.
• LLM-guided yet numerically grounded forecasting: uses structured prompting + confidence-aware gating to inject reliable semantic signals as prefix tokens into a Transformer forecaster, yielding robust gains under event-driven non-stationarity (up to 29% error reduction).

Phase Transition in Nonparametric Minimax Rates for Covariate Shifts on Approximate Manifolds · Code · Poster · Slides
Boston University; Chicago Booth

• New minimax theory for “near-manifold” shift: exposes a sharp phase transition controlled by the support gap between target and source neighborhoods—unifying multiple geometric-transfer regimes.
• Ratio-free, adaptive estimator: achieves near-optimal, dimension-adaptive rates without density ratios and without assuming knowing the geometry (works under approximate manifold mismatch).

Cross-Domain Hyperspectral Image Classification via Mamba-CNN and Knowledge Distillation · Slides
IEEE TGRS 2025

• Hybrid spectral–spatial modeling for domain shift: integrates a Mamba-based global spectral encoder with CNN local feature extraction, capturing long-range dependencies while preserving fine-grained spatial structure in hyperspectral images.
• Dual-level transfer via distillation + graph alignment: performs teacher–student knowledge distillation for distribution alignment and OT-guided graph consistency across domains, yielding robust cross-domain generalization under severe spectral mismatch.

INCM: Inconsistency-Aware Synergistic Multimodal Recommendation via Cross-Modal Hard Negative Sampling
Boston University; Shandong University; Tsinghua University

• Disentangles synergy vs redundancy in multimodal signals: explicitly decomposes item information into unique (text / vision) and synergistic components using heterogeneous–homogeneity graph transformations, avoiding naïve feature fusion.
• Inconsistency-aware learning with principled negatives: introduces cross-modal hard negative sampling and a Synergy-BPR objective that suppresses redundant cues while amplifying truly complementary signals, yielding robust gains under noisy or conflicting modalities.

Semantic Scientific Graph Pruning for Reliable Agentic Paper Reproduction
Boston University; Shandong University

• Semantic pruning for controllable agent search: proposes Semantic Scientific Graph Pruning (SSGP) to transform dense scientific graphs into task-adaptive local subgraphs, using rank-based ensemble scoring to retain only semantically essential neighbors.
• Enables reliable agentic reproduction: couples pruned graphs with reuse–patch modular execution and confidence-weighted aggregation, dramatically reducing search space while improving reproducibility, stability, and success rate of LLM-based scientific agents.