eVQ-AM: An Extended Dynamic Version of Evolving Vector Quantization
Sprache des Vortragstitels:
IEEE SSCI 2013 Conference
Sprache des Tagungstitel:
In this paper, we are presenting a new dynamically
evolving clustering approach which extends conventional
evolving Vector Quantization (eVQ), successfully applied before
as fast learning engine for evolving cluster models, classifiers
and evolving fuzzy systems in various real-world applications.
The first extension concerns the ability to extract ellipsoidal
prototype-based clusters in arbitrary position, thus increasing its
flexibility to model any orentiation/rotation of local data clouds.
The second extension includes a single-pass merging strategy
in order to resolve unnecessary overlaps or to dynamically
compensate inappropriately chosen learning parameters (which
may lead to over-clustering effects). The new approach, termed as
eVQ-AM (eVQ for Arbitrary ellipsoids with Merging functionality),
is compared with conventional eVQ, other incremental and batch
learning clustering methods based on two-dimensional as well
as high-dimensional streaming clustering showing an evolving
behavior in terms of adding/joining clusters as well as feature
range expansions. The comparison includes a sensitivity analysis
on the learning parameters and observations of finally achieved
cluster partition qualities.