Locally
Linear Embedding and its Variants for Feature Extraction
Róbert
Busa-Fekete, András Kocsor
Research
Group on Artificial Intelligence of the Hungarian Academy of Sciences
and University of Szeged
Many problems in machine learning
are hard to manage without applying some pre-processing or feature extraction
method. Two popular forms of dimensionality reduction are the methods
of principal component analysis (PCA)
[2]
and multidimensional scaling (MDS)
[18]
. In this paper we examine Locally
Linear Embedding (LLE), which is an unsupervised, non-linear dimension
reduction method that was originally proposed for visualisation. We
will show that LLE is capable of feature extraction if we choose the
right parameter values. In addition, we extend the original algorithm
for more efficient classification. Afterwards we apply the methods to
several databases that are available at the UCI repository, and then
show that there is a significant improvement in classification performance.