Sampling-enabled scalable manifold learning unveils the discriminative cluster structure of high-dimensional data (SUDE)

We propose a scalable manifold learning (SUDE) method that can cope with large-scale and high-dimensional data in an efficient manner. It starts by seeking a set of landmarks to construct the low-dimensional skeleton of the entire data, and then incorporates the non-landmarks into this skeleton based on the constrained locally linear embedding. This toolkit includes the main code of SUDE, and also two applications for preprocess scRNA-seq and ECG data. This paper has been published in Nature Machine Intelligence, and more details can be seen https://www.nature.com/articles/s42256-025-01112-9.

How To Run

MATLAB

MATLAB code of SUDE is in the sude_mat file, where the sude function provides multiple hyperparameters for user configuration as follows

function [Y, id_samp, para] = sude(X, varargin)
%   This function returns representation of the N by D matrix X in the lower-dimensional space and 
%   the ID of landmarks sampled by PPS. Each row in X represents an observation.
% 
%   Parameters are: 
% 
%   'NumDimensions'- A positive integer specifying the number of dimension of the representation Y. 
%                    Default: 2
%   'NumNeighbors' - A non-negative integer specifying the number of nearest neighbors for PPS to 
%                    sample landmarks. It must be smaller than N.
%                    Default: adaptive
%   'Normalize'    - Logical scalar. If true, normalize X using min-max normalization. If features in 
%                    X are on different scales, 'Normalize' should be set to true because the learning 
%                    process is based on nearest neighbors and features with large scales can override 
%                    the contribution of features with small scales. 
%                    Default: True
%   'LargeData'    - Logical scalar. If true, the data can be split into multiple blocks to avoid the problem 
%                    of memory overflow, and the gradient can be computed block by block using 'learning_l' function.                    
%                    Default: False
%   'InitMethod'   - A string specifying the method for initializing Y before manifold learning. 
%       'le'       - Laplacian eigenmaps.
%       'pca'      - Principal component analysis.
%       'mds'      - Multidimensional scaling.
%                    Default: 'le' 
%   'AggCoef'      - A positive scalar specifying the aggregation coefficient. 
%                    Default: 1.2
%   'MaxEpoch'     - Maximum number of epochs to take. 
%                    Default: 50

The main.m file provides an example

% Input data
clear;
data = csvread('benchmarks/rice.csv');
% data = textread('G:\MATLAB Drive\MATLAB\MNIST\iris.txt');

% Obtain data size and true annotations
[~, m] = size(data);
ref = data(:, m);
X = data(:, 1:m-1);
clear data

% Perform SUDE embedding
t1 = clock;
[Y, idx, para] = sude(X,'NumNeighbors',10);
t2 = clock;
disp(['Elapsed time:', num2str(etime(t2,t1)),'s']);
plotcluster2(Y, ref);

Python

Installation

Supported python versions are 3.8 and above.

This project has been uploaded to PyPI, supporting direct download and installation from pypi

pip install sude

Manual Installation

git clone https://github.com/ZPGuiGroupWhu/SUDE-pkg.git
cd SUDE-pkg
pip install -e .

The SUDE algorithm package provides the sude function for dimension reduction.

The description of the hyperparameters for user configuration are presented as follows

def sude(
    X,
    no_dims = 2,
    k1 = 20,
    normalize = True,
    large = False,
    initialize = 'le',
    agg_coef = 1.2,
    T_epoch = 50,
):
"""
    This function returns representation of the N by D matrix X in the lower-dimensional space. Each row in X
    represents an observation.

    Parameters are:

    'no_dims'      - A positive integer specifying the number of dimension of the representation Y.
                   Default: 2
    'k1'           - A non-negative integer specifying the number of nearest neighbors for PPS to
                   sample landmarks. It must be smaller than N.
                   Default: adaptive
    'normalize'    - Logical scalar. If true, normalize X using min-max normalization. If features in
                   X are on different scales, 'Normalize' should be set to true because the learning
                   process is based on nearest neighbors and features with large scales can override
                   the contribution of features with small scales.
                   Default: True
    'large'        - Logical scalar. If true, the data can be split into multiple blocks to avoid the problem
                   of memory overflow, and the gradient can be computed block by block using 'learning_l' function.
                   Default: False
    'initialize'   - A string specifying the method for initializing Y before manifold learning.
        'le'       - Laplacian eigenmaps.
        'pca'      - Principal component analysis.
        'mds'      - Multidimensional scaling.
                   Default: 'le'
    'agg_coef'     - A positive scalar specifying the aggregation coefficient.
                   Default: 1.2
    'T_epoch'      - Maximum number of epochs to take.
                   Default: 50
"""

After installing the library, you can use the sude function as follows:

import pandas as pd
import numpy as np
from sude import sude
import time
import matplotlib.pyplot as plt

# Input data
data = np.array(pd.read_csv('benchmarks/rice.csv', header=None))

# Obtain data size and true annotations
m = data.shape[1]
X = data[:, :m - 1]
ref = data[:, m - 1]

# Perform SUDE embedding
start_time = time.time()
Y = sude(X, k1=10)
end_time = time.time()
print("Elapsed time:", end_time - start_time, 's')

plt.scatter(Y[:, 0], Y[:, 1], c=ref, cmap='tab10', s=4)
plt.show()

Depends

scRNA-seq application

argparse (≥2.0.4), assertthat (≥0.2.1), BiocGenerics (≥0.40.0), BiocSingular (≥1.10.0), ClusterR (≥1.2.5), dotCall64 (≥1.0.1), fields (≥12.5), GenomeInfoDb (≥1.30.1), GenomicRanges (≥1.46.1), geometry (≥0.4.5), ggplot2 (≥3.3.5), grid (≥4.1.0), gtools (≥3.9.2), IRanges (≥2.28.0), MatrixGenerics (≥1.6.0), mclust (≥5.4.7), parallel (≥4.1.0), prodlim (≥2019.11.13), RcppHungarian (≥0.1), readr (≥1.4.0), reshape2 (≥1.4.4), S4Vectors (≥0.30.0), scran (≥1.22.1), scuttle (≥1.4.0), Seurat (≥4.0.5), SingleCellExperiment (≥1.16.0), spam (≥2.7.0), stats4 (≥4.1.0), SummarizedExperiment (≥1.24.0), uwot (≥0.1.10)

Noted: all R packages can be installed from the CRAN repository or Bioconductor. You can also use the following R scripts to install them all.

## Please click Tools->Global Options->Packages, change CRAN repository to a near mirror. Then, execute the following code:
## Install packages from CRAN.
install.packages(c("argparse", "assertthat", "ClusterR", "dotCall64", "fields", "geometry", "ggplot2", "gtools", "mclust", "prodlim", "RcppHungarian", "readr", "reshape2", "Seurat", "spam", "uwot"))
## Determine whether the package "BiocManager" exists, if not, install this package.
if (!require("BiocManager", quietly = TRUE))
    install.packages("BiocManager")
## Install packages from Bioconductor.
BiocManager::install(c("BiocGenerics", "BiocSingular", "GenomeInfoDb", "GenomicRanges", "IRanges", "MatrixGenerics", "S4Vectors", "scran", "scuttle", "SingleCellExperiment", "SummarizedExperiment"), force = TRUE, update = TRUE, ask = FALSE)

ECG application

Deep Learning Toolbox

Signal Processing Toolbox

Citation Request

Peng, D., Gui, Z., Wei, W. et al. Sampling-enabled scalable manifold learning unveils the discriminative cluster structure of high-dimensional data. Nat. Mach. Intell. (2025). https://doi.org/10.1038/s42256-025-01112-9

Sude

Install / Use

README