Views: 0 | Downloads: 0
Classiers obtained using a machine-learning algorithm can be evaluated by several criteria.
This thesis considers two criteria that are usually treated as the most important when
deciding which classier to apply in practice: comprehensibility (i.e., understandability or
interpretability) and accuracy. While it is well understood which methods and metrics
for estimating a classier's accuracy should be used in a given application, it is more
challenging to measure comprehensibility because the literature on the topic is scarce.
The rst part of this thesis is focused on the comprehensibility of classication trees.
It systematically investigates the inuence of tree-structure parameters (the number of
leaves, branching factor, tree depth) and visualisation properties on the tree comprehensibility.
In addition, we analyse the inuence of the question-depth (the depth of the
deepest leaf that is required when answering a question about a classication tree), which
turns out to be the most important while generally overlooked parameter inuencing
the comprehensibility. The analysis is based on empiric data obtained using a carefully
designed survey with 98 questions answered by 69 respondents grouped according to their
data-mining background. We evaluate several tree-comprehensibility metrics and propose
two new metrics derived from the structure of a tree and user-provided weights, which
enable considering the semantics of the tree in addition to the tree structure itself.
The second problem is that most of the existing machine-learning algorithms are focused
primarily on the classier accuracy, while the ones that take comprehensibility into account
as well, usually treat it as less important. Therefore, in many classication domains the
user is forced to choose between a comprehensible classier with relatively low accuracy or
an incomprehensible but relatively accurate classier.
The second part of the thesis deals with the problem of nding classiers that are
accurate and comprehensible at the same time by proposing a multi-objective machinelearning
algorithm, which treats both objectives as equally important and returns the
set of non-dominated classiers. We propose a hybrid tree as a classication model, which
provides a well-covered range of trade-os between the two conicting objectives. A hybrid
tree is similar to a classication tree, with an exception that some of its leaves use a
black-box classier, which enables trading comprehensibility for accuracy. A corresponding
hybrid-tree comprehensibility metric based on the results from the rst part of the thesis
is also proposed. The proposed algorithm for multi-objective learning of hybrid classiers
(MOLHC) can examine considerably larger hybrid-trees than a naive exhaustive search
algorithm, which faces combinatorial explosion. We show that the MOLHC algorithm is
preferable to general multi-objective optimisation algorithms, because it does not require
any parameter settings (which are dicult to set for the optimisation algorithm), is formally
proven to nd the complete set of non-dominated hybrid trees (while the optimisation
algorithm only tries to converge to it) and has shorter run-times for reasonably sized trees.
A case study in the activity-recognition domain shows that the MOLHC algorithm oers
new insights into the classication problem and learns classiers with a good trade-o
between accuracy and comprehensibility.