Related Experiment Videos
Unsupervised feature evaluation: a neuro-fuzzy approach
1Machine Intelligence Unit, Indian Statistical Institute, Calcutta, 700035, India. sankar@isical.ac.in
IEEE Transactions on Neural Networks
|February 6, 2008
Summary
This study introduces novel neuro-fuzzy methods for unsupervised feature selection and extraction. These approaches identify optimal features and reduce dimensionality without prior cluster knowledge.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Data Science
Background:
- Unsupervised learning requires effective feature selection and extraction for pattern recognition.
- Existing methods often need prior knowledge of cluster numbers or lack flexibility.
Purpose of the Study:
- To develop novel neuro-fuzzy networks for unsupervised feature selection and extraction.
- To introduce a flexible fuzzy feature evaluation index and membership function.
- To perform feature selection and extraction without needing the number of clusters.
Main Methods:
- Formulation of neuro-fuzzy approaches for feature selection and extraction.
- Definition of a fuzzy feature evaluation index based on pattern similarity.
- Introduction of a flexible membership function incorporating weighted distance.
- Design of two new layered networks for unsupervised learning.
Main Results:
- A network for feature selection provides an optimal order of individual feature importance.
- A network for feature extraction identifies optimum transformed features and their relative importance.
- Dimensionality reduction from n-dimensional to n'-dimensional space (n' < n).
- Experimental validation demonstrates superiority over related methods.
Conclusions:
- The proposed neuro-fuzzy networks effectively perform feature selection and extraction in unsupervised learning.
- These methods are robust and do not require prior information on the number of clusters.
- The approach offers a flexible and powerful tool for data analysis and dimensionality reduction.
