Python implementation for Multiple Marginal Fisher Analysis (TIE).

Introduction

MMFA is a supervised subspace learning method. Unlike the most existing methods, MMFA can automatically estimate the feature dimension and obtain the low-dimensional representation.

Affinity Graph Construction

Requirements

• Python 3.7
• numpy
• scikit-learn

Datasets

Here we provide two used datasets in our experiments: AR face images and Extend Yale B face image. We resize the AR images to 55x40 and Yale images to 54x48 size.

Training and Evaluation

python run.py

This should give the classification accuracy results on the AR and Yale datasets.

Or you can simply use MMFA as a python module and perform it on the custom data:

import numpy as np
import mmfa

data, labels = load_data()

# specify k_1, k_2, binary_weight
mapping = mmfa.MMFA(data, labels, k_1, k_2, binary_weight)

low_dimensional_data = np.dot(data, mapping)

# do something with the processed data
...

Citation

If MMFA is useful for your research, please cite the following paper:

@article{huang2018mmfa,
  title = {Multiple Marginal Fisher Analysis},
  author = {Huang, Zhenyu and Zhu, Hongyuan and Zhou, Joey Tianyi and Peng, Xi},
  journal = {IEEE Transactions on Industrial Electronics},
  year = {2018},
  issn = {0278-0046},
  month = dec,
  volume = {66},
  number = {12},
  pages = {9798-9807},
  publisher = {IEEE},
  doi = {10.1109/TIE.2018.2870413},
  html = {https://ieeexplore.ieee.org/document/8476585},
  abbr = {TIE},
  bibtex_show = {true},
  keywords = {Dimensionality reduction;Learning systems;Manifolds;Task analysis;Robustness;Gaussian distribution;Estimation;Automatic dimension reduction;graph embedding;manifold learning;supervised subspace learning}
}

License

MIT License