Summary: | 碩士 === 國立臺灣大學 === 資訊工程學研究所 === 94 === Multi-label classification is an important subject in machine learning. There
are several available ways to handle such problems. In this thesis we focus on using
support vector machines (SVMs). As multi-label classification can be treated
as an extension of multi-class classification, it is natural to modify multi-class approaches
for multi-label problems. The thesis considers three extensions: “binary,”
“label combination” and maximal margin formulation. We give comprehensive experiments
to check their performances. In addition, we also give detailed derivations
and investigate the implementation details.
As “label combination” is a way that treats each subset of labels as an individual
SVM class, any multi-class method can be directly applied. We discuss several methods
of this type. They are “one-against-one,” “approach in [45, 46],” and “method
by Crammer and Singer.” We compare and analyze their performances. The last
two methods both solve a single optimization problem in training. We find that they
perform well when the size of the data is not large. They are however not suitable
for very large problems due to lengthy training time. In such situations, the “label
combination” approach via “one-against-one” multi-class implementation is an
effective solution. Overall we find that the method “label combination” to directly
transform multi-label to multi-class is a practically viable technique.
|