Sparsification Subsampling for Partial Least Squares Regression
TL;DR: Uses one-sided element sparsification to approximate covariance structure and accelerate large-scale PLS via rewritten kernel updates.
BDAL-RUC Research Map
Our work connects scalable statistical computation, representative design, Hilbert and space-filling transport metrics, optimal transport geometry, cross-domain soft alignment, and statistical methods for industrial and AI systems. We aim to build reusable methodology that is both theoretically interpretable and computationally effective.
34Publications listed
1 / 3 / 15RUC A+ / A / A- journals
5NeurIPS / ICML / ICLR
Representative Work SparSink HCPD POTD Core-Elements SEINT MP-MoE
Quick Navigation
Subsampling, Sparsification, and Representative Design
From sample-level subsampling to element-, operator-, and interaction-level sparsification.
Hilbert and Space-Filling Transport MetricsUsing low-dimensional order, mass coordinates, and original-space costs to preserve geometry efficiently.
OT Geometry, Dimension Reduction, and Invariant RepresentationUsing OT as a geometric proxy for discriminative directions and invariant representations.
Cross-Domain Soft Alignment and Applied OTUsing OT and UOT to build soft correspondences for cross-domain inference, registration, and representative objects.
Statistics for Industrial and AI SystemsTurning statistical computation and functional-unit selection into practical algorithms for imaging, deep learning, and MoE systems.
Background ReviewsMapping the technical landscape of subsampling, projection OT, and OT sparsification.
Other Papers and Low-Weight ComplementsCollecting interval-valued objects, applied collaborations, and related works outside the core themes.
Papers by Research Theme
Subsampling, Sparsification, and Representative Design
This line studies representative design, subsampling, and element- or interaction-level sparsification to reduce large-scale computation while preserving statistical properties.
Core-Elements Subsampling for Alternating Least Squares
TL;DR: Embeds element-level subsampling into alternating least-squares updates to accelerate matrix factorization with error control.
Sampling-Based Methods for Multi-Block Optimization Problems over Transport Polytopes
TL;DR: Develops sampling-based solvers for multi-block transport polytopes to reduce the cost of multi-marginal structural optimization.
Nonparametric Additive Models for Billion Observations
TL;DR: Extends element-level sparsification to billion-scale additive models with matching penalized estimation and tuning criteria.
Core-Elements for Large-Scale Least Squares Estimation
TL;DR: Moves subsampling from rows to matrix elements and rewrites the estimator to preserve large-sample properties.
An Optimal Transport Approach for Selecting a Representative Subsample with Application in Efficient Kernel Density Estimation
TL;DR: Uses optimal transport to select representative subsamples that improve distributional coverage for kernel density estimation.
Importance Sparsification for Sinkhorn Algorithm
TL;DR: Derives computable sampling upper bounds from OT structure to sparsify pairwise interactions in Sinkhorn.
Efficient Approximation of Gromov-Wasserstein Distance Using Importance Sparsification
TL;DR: Extends SparSink-style importance sparsification to the computation of Gromov-Wasserstein distances.
LowCon: A Design-Based Subsampling Approach in a Misspecified Linear Model
TL;DR: Uses space-filling subsampling to improve robustness and coverage under model misspecification.
More Efficient Computation of Smoothing Splines via Space-Filling Basis Selection
TL;DR: Uses low-discrepancy space-filling basis selection to accelerate smoothing spline approximation.
Hilbert and Space-Filling Transport Metrics
This line uses Hilbert and space-filling orders to build lightweight structures for distribution comparison and high-dimensional geometric computation.
Efficient Variants of Wasserstein Distance in Hyperbolic Space via Space-Filling Curve Projection
TL;DR: Uses space-filling curves to generate lightweight couplings in hyperbolic space while evaluating original geometric costs.
Hilbert Curve Projection Distance for Distribution Comparison
TL;DR: Uses Hilbert ordering to generate transport couplings, then evaluates costs in the original space for distribution comparison.
Smoothing Splines Approximation Using Hilbert Curve Basis Selection
TL;DR: Uses Hilbert-curve ordering for basis selection, balancing coverage and efficiency in multivariate smoothing splines.
OT Geometry, Dimension Reduction, and Invariant Representation
This line treats optimal transport as a geometric proxy for direction recovery, projection pursuit, feature screening, and rigid-motion invariant metrics.
SSP-Ensemble: A Sufficient Subspace Projection Ensemble for Multiclass Classification
TL;DR: Builds supervised pair-specific sufficient projections for multiclass classification instead of relying on random subspace ensembles.
KPOTD: Kernel Principal Optimal Transport Directions for Nonlinear Sufficient Dimension Reduction
TL;DR: Lifts POTD displacement covariance into an RKHS to recover nonlinear sufficient dimension reduction structures.
An Efficient SE(p)-Invariant Transport Metric Driven by Polar Transport Discrepancy-Based Representation
TL;DR: Constructs information-preserving rigid-motion invariant representations for low-cost transport metrics.
Scalable Model-Free Feature Screening via Sliced-Wasserstein Dependency
TL;DR: Uses sliced-Wasserstein dependency for model-free feature screening and nonlinear variable importance.
Sufficient Dimension Reduction for Classification Using Principal Optimal Transport Direction
TL;DR: Uses OT displacements between class distributions as proxies for recovering sufficient dimension reduction directions.
Large-Scale Optimal Transport Map Estimation Using Projection Pursuit
TL;DR: Uses residual-driven projection pursuit directions to build high-dimensional OT map approximations step by step.
Cross-Domain Soft Alignment and Applied OT
This line uses OT, UOT, and partial OT to create soft correspondences or representative objects for unpaired inference, image registration, and 3D representation compression.
Iterative Optimal Transport for Multimodal Image Registration
TL;DR: Alternates transformation estimation with unbalanced OT soft alignment for multimodal image registration.
Double Optimal Transport for Differential Gene Regulatory Network Inference with Unpaired Samples
TL;DR: Uses two layers of OT to create soft correspondences for differential gene regulatory network inference with unpaired samples.
Gaussian Herding Across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS
TL;DR: Compresses 3D Gaussian representations from an OT perspective by reducing redundant Gaussian components.
Statistics for Industrial and AI Systems
This line translates statistical computation, leverage approximation, active learning, and functional-unit selection into deployable algorithms for real applications.
Breaking the Echo Chamber: A Dynamic Ensemble Pruning Perspective on MoE
TL;DR: Views MoE expert routing as functional-unit pruning to select strong and complementary experts and reduce echo chambers.
SPOT: An Active Learning Algorithm for Efficient Deep Neural Network Training
TL;DR: Uses active learning to select high-value samples and improve deep neural network training efficiency.
Efficient Approximation of Leverage Scores in Two-Dimensional Autoregressive Models with Application to Image Anomaly Detection
TL;DR: Uses two-dimensional autoregressive structure to approximate leverage scores for efficient image anomaly detection.
Background Reviews
These papers serve as background entry points for large-scale subsampling, projection-based optimal transport, and optimal transport sparsification.
Sparsification Techniques for Large-Scale Optimal Transport Problems
TL;DR: Reviews large-scale OT sparsification methods as a background entry point for the SparSink line.
Projection‐Based Techniques for High‐Dimensional Optimal Transport Problems
TL;DR: Reviews projection-based high-dimensional OT techniques as a background entry point for sliced and projection OT.
Modern Subsampling Methods for Large-Scale Least Squares Regression
TL;DR: Reviews large-scale least-squares subsampling methods as a background entry point for leverage and optimal subsampling.
Effective Statistical Methods for Big Data Analytics
TL;DR: Reviews statistical modeling and computation tools for big data analytics as an early background entry point.
Other Papers and Low-Weight Complements
These works are related to applied collaborations or complementary geometric-statistical settings, but are not forced into the core themes above.
Hausdorff Correlation for Interval-Valued Random Objects
TL;DR: Constructs correlation measures for interval-valued random objects as a geometric dependence-analysis tool.
Leverage Classifier: Another Look at Support Vector Machine
TL;DR: Revisits support vector machine classification from a leverage-score perspective.
AppleQSM: Geometry-Based 3D Characterization of Apple Tree Architecture in Orchards
TL;DR: Builds geometry-based characterization and quantification methods for 3D apple-tree architecture in orchards.
A Hausdorff Regression Paradigm for Interval Privacy
TL;DR: Builds a Hausdorff regression framework for interval-valued privacy data.
High-Resolution 3D Terrestrial LiDAR for Cotton Plant Main Stalk and Node Detection
TL;DR: Applies high-resolution 3D LiDAR to characterize cotton main stalks and nodes.