A Novel Unsupervised Feature Selection for High-Dimensional Data Based on FCM and k-Nearest Neighbor Rough Sets

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• High-dimensional unlabeled data often contains limited useful information.
• Existing unsupervised feature selection methods struggle with uneven data densities and high computational costs.

Purpose of the Study:

• To propose an effective unsupervised feature selection technique for high-dimensional data.
• To address limitations of current methods in handling uneven data density and computational time.

Main Methods:

• A novel feature extraction technique combining Fuzzy C-Means (FCM) clustering and k-nearest neighbor rough sets.
• FCM is utilized for clustering, followed by feature importance evaluation and sorting.
• K-nearest neighbor rough sets are employed for filtering redundant features.

Main Results:

• The proposed KND-UFS algorithm demonstrated superior performance across 12 public datasets.
• Significant reductions in running time and effective feature selection were observed.
• KND-UFS outperformed eight existing algorithms in classification accuracy and efficiency.

Conclusions:

• The KND-UFS algorithm offers an efficient and accurate solution for unsupervised feature selection.
• The combination of FCM and k-nearest neighbor rough sets effectively handles datasets with uneven density.
• This method provides a valuable tool for extracting meaningful information from large, unlabeled datasets.