-
In multilabel classi cation, a data sample is classi ed into one class or plural
classes [
1
]. One of the widely used classi cation methods uses one-against-all
classi cation, in which for an n-class problem, n decision functions are
determined, with each decision function putting one class on the positive side and
the remaining classes on the negative side. In classi cation, a data sample is
classi ed into a single-label or multilabel class associated with positive decision
functions. By this method, a data sample is unclassi able if there is no positive
decision function, and a data sample may be classi ed into a multilabel that is
not included in the multilabels contained in the training set.
To solve this problem, in this paper, we propose one-against-all fuzzy support
vector machines (FSVMs) for multilabel classi cation [
2
]. For each multilabel in
the training data set, we de ne a new multilabel class. And for each single label or
multilabel class, we de ne a fuzzy region using the decision functions determined
by one-against-all classi cation. The degree of membership of a data sample to
the fuzzy region is determined by the decision hyperplane that is nearest to the
data sample. And the data sample is classi ed into the class with the highest
degree of membership.
This classi cation strategy is simpli ed for an unclassi able region. If no
decision function is positive for a data sample, it is classi ed into a class with
the maximum degree of membership. This is the same as the fuzzy SVM for
single-class classi cation.
We compare the accuracies and subset accuracies of the proposed FSVMs
with the conventional one-against-all, one-against-one, and the best accuracies
in [
1
] using several benchmark data sets that are used in [
1
].
This work was supported by JSPS KAKENHI Grant Number 25420438.