An enhanced dimensionality reduction for multi-label learning

Locally Linear Embedding (LLE) manifold algorithm is a kind of unsupervised dimensionality reduction algorithm, which preserves geometric structure of the local patches when high-dimensional data is mapped to a low-dimensional space. To solve `curse of dimensionality' problem and enhance the ac...

Full description

Saved in:
Bibliographic Details
Published in2015 International Conference on Computers, Communications, and Systems (ICCCS) pp. 163 - 170
Main Authors Yanqing Shao, Kai Yan
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.11.2015
Subjects
Classification algorithms
Computers
Dimensionality reduction
Electronic mail
locally linear Embedding
Manifolds
Measurement
multi-label learning
Principal component analysis
Smoothing methods
variable K-nearest semi-supervised
Online AccessGet full text

Cover

More Information
Summary:Locally Linear Embedding (LLE) manifold algorithm is a kind of unsupervised dimensionality reduction algorithm, which preserves geometric structure of the local patches when high-dimensional data is mapped to a low-dimensional space. To solve `curse of dimensionality' problem and enhance the accuracy of multi-label learning, Variable K-nearest Locally Linear Embedding (VKLLE) and Variable K-nearest Semi-Supervised Locally Linear Embedding (VKSSLLE) are proposed. Different from existing approaches, the proposed algorithms calculate the proper number of nearest neighbors according to the distribution of sample points in dataset, which avoids smoothing or elimination of small-scale structure in the manifold as well as false division of the continuous manifold into irrelevant sub-manifolds. To verify the performances of the proposed algorithms, different dimensionality reduction algorithms are respectively incorporated with multi-label naive Bayes classifier to solve multi-label learning problem. Results on artificial dataset and two real-world datasets show that proposed algorithms can effectively enhance the accuracy of multi-label learning.
DOI:10.1109/CCOMS.2015.7562894