Locally Weighted Ensemble Clustering

Overview

This repository provides the MATLAB code for two ensemble clustering algorithms, namely, locally weighted evidence accumulation (LWEA) and locally weighted graph partitioning (LWGP), which are proposed in the following paper:

Dong Huang, Chang-Dong Wang, and Jian-Huang Lai.
Locally Weighted Ensemble Clustering,
IEEE Transactions on Cybernetics, 2018, 48(5), pp.1460-1473.

Description of Files

There are mainly two types of files in this repository:

1. A pool of 100 base clusterings for each dataset;
2. The code of the LWEA and LWGP algorithms.

Data

The base clustering pools for the 15 datasets are provided in the following MAT files:

dataName = 'VS';
dataName = 'Semeion';
dataName = 'SPF';
dataName = 'MF';
dataName = 'IS';
dataName = 'Caltech20';
dataName = 'FCT';
dataName = 'MNIST';
dataName = 'Texture';
dataName = 'ODR';
dataName = 'LS';
dataName = 'ISOLET';
dataName = 'PD';
dataName = 'USPS';
dataName = 'LR';

There are two variables in the MAT file for each dataset, namely, members and gt. The variable gt is the ground-truth label, which is an N-dimension vector. The variable members is an N x s matrix, where each column of it is a candidate base clustering.

Code

The file entitled 'demo_LWEA_and_LWGP.m' is the main file for running LWEA and LWGP. You may change the following settings in order to test the performances of LWEA and LWGP:

1) dataName:	the dataset to be used.
2) M:           the ensemble size.
3) cntTimes:	run LWEA and LWGP for cntTimes times and obtain the average performance.
4) para_theta:	the parameter theta.
5) clsNums:     a vector of positve integers, specifying different numbers of clusters for LWEA and LWGP.

(Please note that M, cntTimes must be positive integers. The parameter para_theta must be greater than zero.)

The execution results and the variable 'bcIdx' will be saved in results_XXX.mat. The bcIdx is a cntTimes x M matrix and stores the information of the ensembles. Each row in bcIdx includes M indices for choosing base clusterings in the pool and thus represents an ensemble of M base clusterings. When comparing our approach to other approaches, please make sure that they use the same base clustering settings, i.e., use the ensembles generated by the same 'bcIdx'.

Questions?

Don't hesitate to contact me if you have any questions regarding this work.
Email: huangdonghere at gmail dot com
Website: https://www.researchgate.net/publication/316681928