Graphics Reference
In-Depth Information
+
not agree with all of its k
l nearest neighbors, where l are all the instances in S
which are at the same distance as the last neighbor of X i .
In addition, MENN works with a prefixed number of pairs ( k , k ). k is employed
as the number of neighbors involved to perform the editing process, and k is
employed to validate the edited set S obtained. The best pair found is employed
as the final reference set (if two or more sets are found as optimal, then both are
employed in the classification of the test instances. A majority rule is used to
decide the output of the classifier in this case).
Nearest Centroid Neighbor Edition (NCNEdit) [ 140 ]—The NCN Editing algo-
rithm applies the NCN classification rule to perform an edition process over the
training set TR . The NCN classification rule can be defined as:
Having defined the NCN scheme, the editing process consists in set S
=
TR an
discard from S every prototype misclassified by the NCN rule.
8.4.2.2 Batch
AllKNN [ 149 ]—All KNN is an extension of ENN. The algorithm, for i
0to k
flags as bad any instance not classified correctly by its i nearest neighbors. When
the loop is completed k times, it removes the instances flagged as bad.
=
Model Class Selection (MoCS) [ 16 ]—Brodley's algorithm for reducing the size
of the training set TR is to keep track of how many times each instance was one of
the k nearest neighbors of another instance, and whether its class matched that of
the instance being classified. If the number of times it was wrong is greater than
the number of times it was correct then it is thrown out.
8.4.3 Hybrid Algorithms
Hybrids methods try to find the smallest subset S which lets keep or even increase
the generalization accuracy in test data. For doing it, it allows the removal of internal
and border points.
8.4.3.1 Incremental
Instance-Based Learning Algorithms Family (IB3) [ 1 ]—A series of instance-
based learning algorithms are presented. IB1 was simply the 1-NN algorithm, used
as a baseline.
IB2
-It starts with S initially empty, and each instance in TR is added to S if it is not
classified correctly by the instances already in S (with the first instance always added).
 
Search WWH ::




Custom Search