Home

Awesome

Non-Graph Data Clustering via O(n) Bipartite Graph Convolution

This repository is our implementation of

Hongyuan Zhang, Jiankun Shi, Rui Zhang, and Xuelong Li, "Non-Graph Data Clustering via O(n) Bipartite Graph Convolution," IEEE Transactions on Pattern Analysis and Machine Intelligence, DOI:10.1109/TPAMI.2022.3231470, 2022.

AnchorGAE attempts to accelarate the unsupervised GNN (e.g., AdaGAE), which could be used to promote the clustering performance, via the classical trick of anchors / landmarks. It leads to a Siamese architecture and a specific graph convolution operation.

It should be emphasized that AnchorGAE is designed for the clustering on non-graph data, where all data points are only represented by $d$-dimension vectors and the graph is not provided as priori. It could be regarded as an GNN extension of scalable graph clustering.

If you have issues, please email:

hyzhang98@gmail.com or henusjk@163.com.

How to Run AnchorGAE

To run the experiment, the name of dataset and parameters need be required. The required configuration is explained at the end.

Name of Dataset

There are six datasets are provided.

Parameter: {datasetName}

--datasetName=usps_all

--datasetName=segment_uni

--datasetName=mnist_all

--datasetName=Isolet

--datasetName=fashionMNIST_full

--datasetName=mnist_test

Parameters

There are three hyperparameters that need to be set.

Parameter1: {AnchorNum, type:int, help='Initialize the number of anchors.'}

Parameter2: {k0, type:int, Initialize the k-sparsity.}

Parameter3: {increase_k, type:int, help='Initialize the size of the self-increasing sparsity.'}

Run

There is an example running on USPS dataset.

python Main.py --datasetName=usps_all --AnchorNum=400 --increase_k=6 --k0=3

Requirements

Citation

@article{AnchorGAE,
  author={Zhang, Hongyuan and Shi, Jiankun and Zhang, Rui and Li, Xuelong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Non-Graph Data Clustering via O(n) Bipartite Graph Convolution}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2022.3231470}
}