ABSTRACT:
This thesis proposes four new methods for
generating ensembles of classifiers and some pruning and ordering
heuristics that improve the performance of ensembles generated with
bagging algorithm:
- Three
new methods to generate
ensembles of classifiers that use all available data to construct
every individual classifier are introduced. The base algorithm,
presented in
Gelfand et al. (S.B. Gelfand, C.S. Ravishankar, y E.J. Delp. An
iterative growing and pruning algorithm for classification tree design.
IEEE Transactions on
Pattern Analysis and Machine Intelligence, 13(2):138--150,
1991), builds a decision tree in an iterative
manner: The training data is divided into two subsets. In each
iteration, one subset is used to grow the decision tree, starting from
the decision tree produced by the previous iteration. This fully grown
tree is then pruned by using the other subset. The roles of the data
subsets are interchanged in every iteration. This process converges to
a final tree which is stable with respect to the sequence of growing
and
pruning steps. To generate a variety of classifiers for the ensemble we
randomly partition the training data into the subsets needed by the
iterative tree construction
algorithm. The method exhibits good performance in several standard
datasets at a low computational cost.
- Ensembles that combine the decisions of classifiers
generated by using perturbed versions of the training set where the
classes of the training examples are randomly switched can produce a
significant error reduction, provided that large numbers of units and
high class switching rates are used. The individual classifiers
generated by this
procedure have statistically uncorrelated errors in the training set.
Hence, the ensembles they form exhibit a similar dependence of the
training error on ensemble size, independently of the classification
problem. In particular, for binary classification problems, the
classification performance of the ensemble on the training data can be
analysed in terms of a Bernoulli process. Experiments on 15 UCI
datasets demonstrate the improvements in classification accuracy that
can be obtained using these class-switching ensembles.
- The order in which classifiers are aggregated in ensemble
methods can be a useful tool for the identification of subsets of
classifiers that, when combined, perform better than the whole
ensemble. Ensembles with randomly ordered classifiers usually
exhibit a generalization error that decreases as the number of
classifiers that are aggregated increases. If an appropriate order for
aggregation is chosen, the generalization error reaches, at
intermediate numbers of classifiers, a minimum, which lies below the
asymptotic error of the ensemble. The second part of this thesis
presents some heuristics that exploit the relations between the
classifiers in a bagging ensemble to identify the appropriate ordering
and then select the first K classifiers for aggregation according to a
desired amount pruning or other pruning rules. The resulting
subensembles are smaller and improve the classification
performance of the original ensemble.
|