Interactive Comprehensible Data Mining - Ambient Intelligence for Scientific Discovery - page 54

comprehensible results. Rules are inherently much more understandable than
decision trees or probabilistic or statistical descriptions.
Thirdly, the genetic algorithm aims to discover rules and rule sets which
optimize an objective function (“fitness”), and manipulation of this allows us
to explore different areas of the search space. For example, we can strongly
penalize rules that give false positive in order to obtain rules that can be used to
determine the class of new data examples. Alternatively, we can bias the system
towards rules that indicate the typical characteristics of items in a group, whether
these characteristics are shared with another group or not. Short rules with few
terms are going to be easier to comprehend than longer ones, but longer rules
reveal more information. We can allow the user to choose which they would
prefer by controlling the fitness function. Initially we might prefer short rules,
in order to get an overview. Note that since the Haiku system is interactive
and iterative, when we have a higher level of comprehension, we can repeat the
process but allow the rules to become longer and hence more detailed, moving
our understanding from a broad overview to comprehending the details.
We use a special type of GA that evolves rules; these produce terms to
describe the underlying data of the form:

IF term OP value | range (& ...) THEN term OP value | range (& ...)

where term is a class from the data set, OP is one of the standard comparison

, £ , ∧ ), value is a numeric or symbolic value, and range is a
numeric range. A typical rule would therefore be:

operators ( <, >, =

colour texture size < fruit

IF =strawberry

There are three situations that are of particular interest to us:

=red & =soft & 3.2 THEN

- Classification: when the left hand side of the equation tries to predict a single

class (usually known) on the right hand side

- Characterization: when the system tries to find rules that describe portions

of the data set

- Association: which detects correlations in attribute values within a portion

of the data set

The algorithm follows fairly typical genetic algorithmic approaches in its
implementation, but with specialized mutation and crossover operators, in order
to explore the space effectively. We start with a number of random rules created
using values from the data. The rules population is then evolved based on how
well they perform. The fittest rules are taken as the basis for the next population,
with crossover creating new rules from clauses of previously successful rules.
Mutation is specialized: for ranges of values it can expand or contract that range;
for numbers it can increase or decrease them; for operators it can substitute them
with others.
Statistically principled comparisons showed that this technique is at least as
good as conventional machine learning at classification [1], but has advantages
over the more conventional approaches in that it can discover characteristics and
associations too.