spectralclustering-1.0
所属分类：数值算法/人工智能
开发工具：matlab
文件大小：15429KB
下载次数：86
上传日期：2009-07-06 22:59:59
上 传 者：
dearliory
说明： 一款功能强大的分类算法软件包，在matlab环境下使用。详见如下网站
http://mloss.org/software/view/133/
(A MATLAB spectral clustering package to deal with large data sets. Our tool can handle large data sets (200,000 RCV1 data) on a 4GB memory general machine. Spectral clustering algorithm has been shown to be more effective in finding clusters than some traditional algorithms such as kmeans. To perform clustering on large data sets, we implement various ways of approximating the dense similarity matrix, including nearest neighbors and the Nystrom method. )
文件列表:
spectralclustering-1.0 (0, 2008-09-04)
spectralclustering-1.0\gen_nn_distance.m (4230, 2008-08-20)
spectralclustering-1.0\sc.m (2876, 2008-08-20)
spectralclustering-1.0\nystrom.m (5000, 2008-08-20)
spectralclustering-1.0\nmi.m (1904, 2008-08-20)
spectralclustering-1.0\k_means.m (4550, 2008-09-02)
spectralclustering-1.0\script_sc.m (2393, 2008-08-20)
spectralclustering-1.0\script_nystrom.m (2095, 2008-08-20)
spectralclustering-1.0\data (0, 2008-08-20)
spectralclustering-1.0\data\corel_100_NN_sym_distance.mat (2626906, 2008-08-20)
spectralclustering-1.0\data\corel_10_NN_sym_distance.mat (258581, 2008-08-20)
spectralclustering-1.0\data\corel_150_NN_sym_distance.mat (3982453, 2008-08-20)
spectralclustering-1.0\data\corel_15_NN_sym_distance.mat (385884, 2008-08-20)
spectralclustering-1.0\data\corel_200_NN_sym_distance.mat (5315835, 2008-08-20)
spectralclustering-1.0\data\corel_20_NN_sym_distance.mat (515644, 2008-08-20)
spectralclustering-1.0\data\corel_50_NN_sym_distance.mat (1297736, 2008-08-20)
spectralclustering-1.0\data\corel_5_NN_sym_distance.mat (129836, 2008-08-20)
spectralclustering-1.0\data\corel_feature.mat (1292947, 2008-08-20)
spectralclustering-1.0\data\corel_label.mat (236, 2008-08-20)
Table of Contents
=================
- Introduction
- Usage
- Examples
- Hardware Requirement
- Additional Information
Introduction
============
This directory includes sources used in the following paper:
Wen-Yen Chen, Yangqiu Song, Hongjie Bai, Chih-Jen Lin, and Edward Chang.
Parallel Spectral Clustering, 2008.
http://www.cs.ucsb.edu/~wychen/publications/PSC08.pdf
A short version appears at ECML/PKDD 2008.
This code has been tested under ***-bit Linux environment using
MATLAB 7.4.0.287 (R2007a).
You will be able to regenerate experiment results in the paper. However,
results may be slightly different due to the randomness, the CPU speed,
and the load of your computer.
In the data/ directory, we include the Corel data set and its nearest
neighbors files. For the RCV1 data set, due to its huge size (~100MB),
please download it at:
RCV1 data : http://www.cs.ucsb.edu/~wychen/download/rcv_feature.mat (***MB)
RCV1 label: http://www.cs.ucsb.edu/~wychen/download/rcv_label.mat (146KB)
To generate nearest neighbors for RCV1, please refer to the Examples section.
Please note that we assume true/predicted labels are integers 1, 2,
..., num_labels.
Usage
=====
1. Generate (t-nearest-neighbor) sparse distance matrix:
matlab> gen_nn_distance(data, num_neighbors, block_size, save_type)
-data:
N-by-D data matrix, where N is the number of data, D is the number of dimensions.
-num_neighbors:
Number of nearest neighbors.
-block_size:
Block size for partitioning the data matrix. We process the data matrix in a
divide-and-conquer manner to alleviate memory use. This is useful for processing
very large data set when physical memory is limited.
-save_type:
0 for .mat file, 1 for .txt file, 2 for both.
[Note] The file format of .txt is as follows:
data_id #_of_neighbors data_id:distance_value data_id:distance_value ...
2. Run spectral clustering using a sparse similarity matrix:
matlab> [predict_labels evd_time kmeans_time total_time] = sc(A, sigma, num_clusters);
-A:
N-by-N sparse symmetric distance matrix, where N is the number of data.
-sigma:
Sigma value used in computing similarity (S), where S_ij = exp(-dist_ij^2 / 2*sigma*sigma);
if sigma is 0, apply self-tunning technique, where S_ij = exp(-dist_ij^2 /2*avg_dist_i*avg_dist_j).
-num_clusters:
Number of clusters.
3. Run sepctral clustering using Nystrom method:
matlab> [predict_labels evd_time kmeans_time total_time] = nystrom(data, num_samples, sigma, num_clusters);
-data:
N-by-D data matrix, where N is the number of data, D is the number of dimensions.
-num_samples:
Number of random samples.
-sigma:
Sigma value used in computing similarity (S), where S = exp(-dist^2 / 2*sigma*sigma).
-num_clusters:
Number of clusters.
4. Run kmeans clustering after eigendecompostion:
matlab> cluster_labels = k_means(data, centers, num_clusters);
-data:
N-by-D data matrix, where N is the number of data, D is the number of dimensions.
-centers:
K-by-D centers matrix, where K is num_clusters, or
'random', random initialization, or
[], empty matrix, orthogonal initialization
-num_clusters:
Number of clusters.
5. Evaludate clustering quality using NMI (Normalized Mutual Information):
matlab> score = nmi(true_labels, predict_labels)
-true_labels:
N-by-1 vector containing true labels.
-predict_labels:
N-by-1 vector containing predicted labels.
6. Run experiment scripts (sparse/nystrom) with different parameters as shown in papers:
matlab> script_sc(dataset)
-dataset:
data set number, 1 = Corel, 2 = RCV1.
matlab> script_nystrom(dataset)
-dataset:
data set number, 1 = Corel, 2 = RCV1.
Examples
========
Ggenerate (t-nearest-neighbor) sparse distance matrix:
matlab> load data/corel_feature.mat;
matlab> gen_nn_distance(feature, 50, 10, 2);
Run spectral clustering using a sparse similarity matrix:
matlab> load data/corel_50_NN_sym_distance.mat;
matlab> [predict_labels evd_time kmeans_time total_time] = sc(A, 20, 18);
matlab> load data/corel_label.mat;
matlab> nmi_score = nmi(label, predict_labels);
Run spectral clustering using Nystrom method:
matlab> load data/corel_feature.mat;
matlab> [predict_labels evd_time kmeans_time total_time] = nystrom(feature, 200, 20, 18);
matlab> load data/corel_label.mat;
matlab> nmi_score = nmi(label, predict_labels);
Run experiment scripts (sparse/nystrom) for Corel data:
matlab> script_sc(1);
matlab> script_nystrom(1);
Hardware Requirement
====================
Please be noted when running Nystrom with RCV1 data (193,844 instances), it may
consume more than 3GB memory with 2000 random samples (193,844 * 2,000 * 8Bytes).
If you want to run with more number of samples, please make sure you have
enough memory on your machine.
Additiona Information
=====================
If you find this tool useful, please cite it as
Wen-Yen Chen, Yangqiu Song, Hongjie Bai, Chih-Jen Lin, and Edward Chang.
Parallel Spectral Clustering, 2008.
For any question, please contact Wen-Yen Chen
and Chih-Jen Lin .
近期下载者：
相关文件：
收藏者：