Information Technology Reference
In-Depth Information
for the choice of k. Weiss et al. provide better solutions in the cases where
feature values are symbols or discrete ones(Weiss, 1991). Aha gives a method,
named as cross validation, to determine the value of k (Aha,1997).
5.10.1
Learning tasks of IBL
In IBL, an instance is described as a set of feature-value pairs. Each instance
contains several features, and missing feature values are tolerated. All
instances are described by the same n features, which define an n-dimensional
instance space. Exactly one of there features corresponds to the category
feature; while the other n-1 features are referred to as predictor features. IBL
algorithms can learn a lot of overlapping concepts, yet in general the learning
only involves exactly one category feature and the categories are disjoint does
not overlap and its outputs are basically simple.
Generally speaking, an output of IBL algorithms is a concept description,
which is a function mapping from instances to categories: Given an instance
in the instance space, the function gives a classification, i.e., a predication on
this instance's category attribute. An instances-based concept description
includes a set of stored instances, and possibly information about their
performances in the past classification processes. This set of instance can
change after each query instance is processed.
Classification function
: It receives as the input the similarity function's
results and the classification performance records for instances, and outputs a
classification for i.
: It maintains records on
classification performance, and determines which instances should be
incorporated in the concept description. IBL assumes that similar instances
have similar classifications, thus it classifies the new instances according to
their most similar neighbors' classifications. Meanwhile, when prior
knowledge is absent, IBL assumes that all features' contributions to
classification are equal, i.e. feature weights reflecting their importance are
identical in the similar function. This bias requires normalizations of each
features' value domain.
Different from most other supervised learning algorithms, no explicit
abstractions such as decision trees or decision inductions are needed in IBL
algorithms. Most learning algorithms maintain a set of generalizations of
instances to form their abstract representations, and adopt simple matching to
Concept description updater
