Inference of Co-occurring Classes: Multi-class and Multi-label Classification - Computational Intelligence Paradigms in Advanced Pattern Classification - page 176

basic emotions [16], such as anger , happiness , fear , sadness , surprise and disgust ,
the dimensional approach refers to dimensions such as active - passive , positive -
negative . In the Mind Reading taxonomy [6], which is an example of the prototypi-
cal approach, comprises 24 meaning groups such as unfriendly , kind , happy,
romantic , thinking and more, each meaning group comprises various affective
states that share a concept or a meaning, such as comprehending , considering ,
choosing , thinking , thoughtful , and more [6] . For multi-disciplinary fields and for
dynamic processes the issue of taxonomic representation is still to be developed.
The manner, in which knowledge is intuitively grasped and organized by hu-
mans is called Schema or mental schema . For example, the manner in which
students generate schemas at the end of a course is a way to assess their under-
standing of the course material and the interconnections between the various
taught subjects, or the manner by which different people annotate data in data re-
positories and pose queries for data search and retrieval. The prototypical tax-
onomic approach most resembles mental schemas. However, for simplicity of the
models, most taxonomic methods associate each item with a single category or
class, and each class at a single location in the defined space. This is a rather sim-
plistic representation of the knowledge domain, which is not always suitable for
the complexity required by real-world applications.
Taxonomic methods are often used in order to reduce the large number of terms
into a manageable and processable knowledge, both for people and for machines.
On the other hand, methods have been devised for dealing with irreducible and
even extendable huge numbers of labels, such as the ever extending personal im-
ages archives, and for all of the documents on the web.
However, ontologies can also bridge the gap between datasets which are based
on different Schemas or taxonomic representations [22]. For scene analysis, i.e.
the existence of independent objects, the term “bag of words” is commonly used
to describe the scene content [7]. Finding statistical relations or co-occurrences be-
tween these objects is one of the explored approaches to multi-label classification
in these cases.
In “real world” applications, the issue of annotation or labeling poses various
problems. Manual labeling is time consuming. In addition, manual annotation of
data is subject to subjective schemas and to human error. This brings forward the
question of reliability and consistency of the “ground truth”, i.e. the training set,
upon which the entire classification is built. This arises when the represented
knowledge is implicit, i.e. is not directly represented, and people or annotators
have to draw on their representations of the knowledge domain in order to anno-
tate. This is the case in most cases of human non-verbal behavior, which is used to
express various types of information and most of the inter-personal communica-
tion, and for mining of scientific knowledge, such as in Geography [50].
Manual labeling [15, 57, 71] comprises either on a relatively large group of hu-
man annotators, either from the wide population, such as using world wide web (In-
ternet) users in various web domains, using a closed set of labels taken from a li-
mited vocabulary set (taxonomy or ontology) or free labeling, or to use a relatively
small group of annotators, mostly experts, to either verify the previous labeling or
to generate the initial set of labels, or rules. Software tools have been devised in