Application-independent feature construction from noisy samples
Résumé
When training classifiers, presence of noise can severely harm their
performance. One may differentiate class noise from attribute noise.
The earlier has been extensively studied while a few methods have
been recently developed to handle the latter. Among the recent
approaches, detecting, cleansing and correcting are the main words.
In this paper, we focus on attribute noise and we consider how a
frequent fault-tolerant (FFT) pattern mining task can be used to
support noise-tolerant classification. Our method copes with noise
without changing or removing any information from the data. It is
based on an application independent strategy for feature
construction based on the so-called delta-free pattern type which
has been proposed earlier as an approximate condensed
representations of frequent itemsets. Our experimental evaluation on
noisy training data sets shows accuracy improvement when using the
computed features instead of the original ones.