SCFS: Subspace Clustering unsupervised Feature Selection

Official implementation of the paper:

Unsupervised Feature Selection Based on Adaptive Similarity Learning and Subspace Clustering

Published in Engineering Applications of Artificial Intelligence (Elsevier), Volume 95, 2020.

Authors:

• Mohsen Ghassemi Parsa
• Hadi Zare
• Mehdi Ghatee

---

Keywords

Unsupervised Feature Selection, Feature Selection, Subspace Clustering, Adaptive Similarity Learning, Representation Learning, Sparse Learning, Dimensionality Reduction, Data Mining, Machine Learning, Clustering, High-Dimensional Data Analysis.

---

If You Use This Code, Please Cite the Paper

This repository accompanies a peer-reviewed journal publication.

If SCFS contributes to your research, experiments, benchmark comparisons, or software, please cite:

@article{Parsa2020SCFS,
  title={Unsupervised Feature Selection Based on Adaptive Similarity Learning and Subspace Clustering},
  author={Parsa, Mohsen Ghassemi and Zare, Hadi and Ghatee, Mehdi},
  journal={Engineering Applications of Artificial Intelligence},
  volume={95},
  pages={103855},
  year={2020},
  publisher={Elsevier},
  doi={10.1016/j.engappai.2020.103855}
}

Paper DOI:

https://doi.org/10.1016/j.engappai.2020.103855

⭐ If this repository helps your work, please consider giving it a star and citing the paper.

---

Method Overview

The proposed SCFS framework jointly performs:

• Subspace Learning
• Cluster Analysis
• Adaptive Similarity Learning
• Sparse Regression
• Feature Selection

Unlike many existing unsupervised feature selection approaches, SCFS learns sample similarities adaptively during optimization instead of relying on a fixed graph constructed beforehand.

Framework

Overall framework of SCFS. The method integrates subspace learning, cluster analysis, and sparse learning into a unified feature selection framework.

---

Illustrative Example

Illustrative example of SCFS. The clustering matrix, similarity matrix, and sparse regression matrix are jointly learned to identify the most informative features.

---

Why SCFS?

Many unsupervised feature selection methods suffer from:

• Fixed similarity graphs
• Two-stage optimization
• Suboptimal graph construction
• Weak integration between clustering and feature selection

SCFS addresses these limitations by introducing a unified optimization framework that jointly learns cluster structure and feature importance.

Main Contributions

✔ Adaptive similarity learning

✔ Implicit similarity matrix construction

✔ Sample-level self-expression model

✔ Joint optimization framework

✔ Sparse feature selection via L2,1 regularization

✔ Convergence-guaranteed optimization algorithm

✔ Extensive experimental validation

---

Key Results

The original paper evaluated SCFS on nine benchmark datasets from biology, computer vision, speech recognition, and text mining.

Performance Highlights

Metric	Result
Best ACC	8 / 9 datasets
Best NMI	Consistently among top performers
Robustness (CV)	Best average performance
Stability	Top-performing methods
Convergence	Fast convergence in practice

Compared Methods

SCFS was compared against:

• LS
• UDFS
• NDFS
• MCFS
• LLCFS
• SPUFS
• LDSSL
• SCUFS
• TraceRatio
• MaxVar

Experimental results demonstrated that SCFS consistently achieves state-of-the-art or highly competitive performance.

---

Paper

Mohsen Ghassemi Parsa, Hadi Zare, Mehdi Ghatee

Unsupervised Feature Selection Based on Adaptive Similarity Learning and Subspace Clustering.

Engineering Applications of Artificial Intelligence, Volume 95, Article 103855, 2020.

DOI:

https://doi.org/10.1016/j.engappai.2020.103855

Publisher:

Elsevier

Journal:

Engineering Applications of Artificial Intelligence (EAAI)

---

Datasets

The original paper evaluates SCFS on the following benchmark datasets:

Dataset	Domain
Lung	Biology
Lymphoma	Biology
Prostate-GE	Biology
ORL	Face Recognition
Isolet	Speech Recognition
BASEHOCK	Text Mining
BA	Image Processing
GLIOMA	Bioinformatics
Madelon	Artificial Dataset

---

Applications

SCFS can be used in:

• Bioinformatics
• Gene Expression Analysis
• Cancer Classification
• Medical Data Mining
• Computer Vision
• Pattern Recognition
• Text Mining
• Document Analysis
• Representation Learning
• Clustering Preprocessing
• High-Dimensional Data Analysis

---

Frequently Asked Questions

Is SCFS supervised?

No. SCFS is an unsupervised feature selection algorithm.

Does SCFS require labels?

No.

Can SCFS be used before clustering?

Yes. This is one of its primary applications.

How are features ranked?

Features are ranked using the L2 norm of rows of the learned transformation matrix W.

Why does SCFS outperform many graph-based methods?

Because similarity learning is integrated into the optimization process rather than computed once beforehand.

---

Citation

Please cite the following paper if you use this repository:

@article{Parsa2020SCFS,
  title={Unsupervised Feature Selection Based on Adaptive Similarity Learning and Subspace Clustering},
  author={Parsa, Mohsen Ghassemi and Zare, Hadi and Ghatee, Mehdi},
  journal={Engineering Applications of Artificial Intelligence},
  volume={95},
  pages={103855},
  year={2020},
  publisher={Elsevier},
  doi={10.1016/j.engappai.2020.103855}
}

---

License

This repository is provided for academic and research purposes.

---

Contact

Mohsen Ghassemi Parsa

Email: mgparsa@ut.ac.ir

GitHub: https://github.com/mohsengh

For questions, bug reports, suggestions, or collaborations, please open an issue or submit a pull request.

---

Support the Project

If you find SCFS useful:

⭐ Star this repository

📄 Cite the paper

🔄 Share the repository with other researchers

These actions help increase the visibility and impact of the research.