Inference of Co-occurring Classes: Multi-class and Multi-label Classification - Computational Intelligence Paradigms in Advanced Pattern Classification - page 186

Similar methods can be seen in music genre and mood annotation. Lukashe-
citch et al. [39], review methods for genre annotation, such as binary SVM-based
classification, and a modified a k-Nearest Neighbors classifier that directly handle
multi-label data. They suggest a method to disambiguate complex and labels like
“world music” which refer in practice to a large variety of genres, using an inter-
mediate abstraction level, they call domains , which are based on sets of features
that represent a certain aspect of the music, such as rhythm and melody.
7.1 Semi-supervised (Annotation) Methods
Semi-automated methods, and semi-supervised methods, such as active learning
[56], have been devised in order to deal with the labeling of large volumes of data
using only a relatively small number of manually labeled samples. Another type of
semi-supervised methods refers to cases in which the samples are only partially
annotated, i.e. each sample is associated with a single label or a few labels, but not
with all the relevant labels [60]. The first type is discussed in this section. An ex-
ample of the second type is presented in the next section.
An example is the multi-class problem of identifying the people in surveillance
video recordings [73, 78]. One method is co-training , in which the relatively small
training set of manually annotated samples is extended automatically and iterative-
ly. At each iteration, the un-annotated samples which are the most similar to the al-
ready annotated samples are given the labels of the nearest classes, respectively,
and added to the training set [2, 73]. Other methods are co-EM and boosting, or
combinations of co-training and boosting. Co-EM does not commit to a label for
the unlabeled data, instead it uses probabilistic labels that may change from one ite-
ration to the other. Co-EM converges as quickly as EM does, but it is applicable to
situations where there are plenty of training data, especially when the feature space
is of high dimensionality [78]. Feng and Xu [18] present a method for semi-
supervised multi-label and multi-instance annotation, in which each label in an im-
age is associated with a different instance or an area of the image, rather than with
the entire image. They review semi-supervised graph-based learning algorithms.
8 Inference of Co-occurring Affective States from Non-verbal
Speech
This section describes a classification method of multi-class and semi-blind multi-
label classification with feature sparsity, which was designed for inferring com-
plex affective states from their non-verbal expressions in speech. The domain of
affective states and human behavior has many common characteristics with other
human knowledge and behavior domains. The requirements of the classification
process, its implementation and the further analysis which was based on it may
provide a relatively simple answer to a wide range of classification problems in
various knowledge domains. More details can be found in [58-60].