Deep Embedded K-means Clustering

Paper

Usage

1. Clone the code locally.

git clone https://github.com/spdj2271/DEKM.git DEKM

2. Launch an experiment on MNIST dataset.

cd DEKM
python DEKM.py MNIST

3. Launch an experiment on other dataset, e.g., 'USPS', 'COIL20', 'FRGC'.

python DEKM.py USPS

• All datasets can be downloaded here.
• When launch experiments on other datasets (except 'MNIST'), you should make sure you have the following folder structure:

|-- undefined
    |-- DEKM.py
    |-- DEKM_dense.py
    |-- utils.py
    |-- datasets
    |   |-- 20NEWS
    |   |   |-- test_data.npz
    |   |   |-- test_label.npz
    |   |   |-- train_data.npz
    |   |   |-- train_label.npz
    |   |-- COIL20
    |   |   |-- COIL20.h5
    |   |-- FRGC
    |   |   |-- FRGC.h5
    |   |-- RCV1
    |   |   |-- test
    |   |   |-- validation
    |   |-- REUTERS
    |   |   |-- 10k_feature.npy
    |   |   |-- 10k_target.npy
    |   |-- USPS
    |       |-- USPS.h5

Result

Average results of three runs:

image (Conv) dataset	ACC	NMI	text (Dense) dataset	ACC	NMI
MNIST	95.75	91.06	REUTES-10K	76.28	59.06
USPS	79.75	82.23	20NEWS	41.08	40.28
COIL-20	69.03	80.06	RCV1-10K	67.15	46.18
FRGC	38.59	50.78

Dependencies

tensorflow 2.4.1

scikit-learn 0.23.2

numpy 1.19.5

scipy 1.2.1

Citations

If you find our project helpful, your citations are highly appreciated:

@inproceedings{guo2021deep,
  title={Deep Embedded K-Means Clustering},
  author={Guo, Wengang and Lin, Kaiyan and Ye, Wei},
  booktitle={2021 International Conference on Data Mining Workshops (ICDMW)},
  pages={686--694},
  year={2021},
  organization={IEEE}
}