DCN: Deep Clustering Network

This repository contains a PyTorch implementation of the paper:

"Towards K-means-friendly Spaces: Simultaneous Deep Learning and Clustering" Jianbo Yang, Devin M. Kaufman, Jinsung Yoon, and Mihaela van der Schaar, original paper

The code is originally from : @xuyxu and @guenthereder

I have merged the code in a jupyter notebook and added some minor changes.

Overview

Deep Clustering Network (DCN) is a method that jointly optimizes a deep autoencoder and K-means clustering objective. The goal is to learn a feature space where K-means performs well, combining representation learning and clustering in a unified framework.

image credits.

This implementation includes:

• A configurable deep autoencoder

• Joint training with K-means loss

• Evaluation metrics: NMI and ARI

• Comparisons with vanilla K-means (on raw data and autoencoder features)

Install requirements

• Scikit-learn:

pip install -U scikit-learn

• Pytorch:

pip install torch torchvision (without CUDA) or

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126 (with CUDA 12.6)

• Pandas and Matplot:

pip install pandas , pip install matplotlib

Experiment

Dataset

The dataset used for the experiments is the mnist dataset:

Pre-training

The reconstruction loss:

Training

The ARI and NMI scores during the training:

Test

The ARI, NMI, ACC scores on the test set:

NMI	ARI	ACC
84.22	75.76	83.34

On the original paper:

NMI	ARI	ACC
81.-	75.-	83.-

The ARI, NMI, ACC scores of the vanilla Kmeans:

NMI	ARI	ACC
43.01	39.89	49.00

Visualisation of the feature latent space