Reciprocal Multi-Layer Subspace Learning for Multi-View Clustering

This is the Matlab implementation of Reciprocal Multi-Layer Subspace Learning for Multi-View Clustering, published in ICCV 2019. Contact: Ruihuang Li ([email protected])

Paper

• We propose a method to hierarchically identify the underlying cluster structure of high-dimensional data by constructing reciprocal multi-layer subspace representations.
• Based on reconstruction, we learn the latent representation by enforcing it to be close to different view-specific subspace representations, which implicitly co-regularizes subspace structures of all views to be consistent to each other.
• With the introduction of neural networks, more general relationships among different views will be explored.

Example Results

Data

In this example, we load ORL and BBCSport datasets. The former contains 400 face images of 40 distinct subjects, from which 3 types of features are extracted. The latter is a collection of 544 documents associated with 2 views taken from sports articles in 5 topical areas.

Run

• First of all, run initial.m to generate a group of view-specific subspace representations as the initialization.

• Second, run demo_FMR.m

Cite

Please cite following papers if you use this code in your own work:

@InProceedings{Li_2019_ICCV,
author = {Li, Ruihuang and Zhang, Changqing and Fu, Huazhu and Peng, Xi and Zhou, Tianyi and Hu, Qinghua},
title = {Reciprocal Multi-Layer Subspace Learning for Multi-View Clustering},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}