Inference of Co-occurring Classes: Multi-class and Multi-label Classification - Computational Intelligence Paradigms in Advanced Pattern Classification - page 173

abilities of specific groups in the population, medical diagnostic, surveillance, ex-
tended speech recognition and synthesis, animation, feedback systems, personaliza-
tion of services, i.e. for adjusting their parameters to individual requirements and
preferences, for authentication of a user identity for security applications, and for
scientific purpose, i.e. for extending the knowledge of physiological, neurological
and cognitive processes, and more. Other applications refer to the analysis of the
behavior of masses which are relevant to sociology and to biology. This can be ex-
tended to the general case of representation, inference or mining, and analysis of
human knowledge.
Another example is the contribution of various scientific fields, with and with-
out practical applications. Multi-class and multi-label classification methods have
also been extensively researched in the field of Bioinformatics , in applications
such as gene function categorization [5, 77]. They are also applicable to fields
such as the analysis of geographical information [50]. Search and recognition of
objects and scene analysis, can be useful for intelligent systems and robots, for
applications that range from classifying products quality as well as for autonom-
ous and semi-autonomous mobile robots and unmanned vehicles, for exploratory
applications and for domestic, public and emergency services which require situa-
tion awareness, retrieval of objects and avoidance of others, and more [1, 62]. Ad-
ditional research fields can benefit from analysis which is based on multi-class and
multi-label classification methods. These could eventually develop to enable the
analysis of complex knowledge domains and processes that have traditionally
been investigated in different fields of social sciences, life sciences, arts and hu-
manities. Examples from the art domains include the analysis of music genre and
mood [39], color retrieval [25], and analysis of aesthetics characteristics and ge-
nres, with applications for research, annotation of art archives and retrieval from
them, and also for applications in fields such as architecture and commercial mul-
ti-media design.
3 The Classification Process
The classification process does not start with the classification, but is rather pre-
ceded by several preliminary stages that may affect the classification results.
These include the definition, recording or selection of the input data, annotation or
labeling, i.e. defining the classes to be found, and giving each data sample one or
more of the class names or class labels, and pre-processing, extraction of features
from the raw data that can be automatically processed.
The first stage of the classification process usually concerns the recording and
gathering of data or the choice between existing datasets. The type and quality of
the gathered information and the definition of the available classes in it dictate, to a
large extent, the capabilities of the classification system (“garbage in - garbage out”
being the undesired case). There are several issues that cannot always be chosen, or
that overcoming them can be achieved at a great cost (time, space, etc), and there-
fore methods have to be devised in order to overcome them. These include the ga-
thering of enough accurate data, accurate annotation or labeling of the data, and the
combination, i.e. gathering enough data to represent each class. Alternatively,