INCREMENTAL CONCEPTUAL CLUSTERING WITHOUT ORDER

                         DEPENDENCY



                        Altintas, Nesip Ilker

               M.S., Department of Computer Engineering

               Supervisor: Assist. Prof. Dr. Cem Bozsahin



                      September 1995, 69 pages



--------------------------------------------------------------------
In this thesis, a new system for incremental conceptual clustering is
presented. Incremental conceptual clustering systems integrate the
learning with performance and obey the basic principles of human
concept learning. As in other incremental learning systems, they are
faced with the problem of dependency on the order of instance
presentation. This study primarily focuses on this problem.

The presented model is capable of learning a hierarchy of concepts
while minimizing the order bias introduced due to its incremental
nature. The learning process is divided into two phases:
nonincremental clustering part and incremental refinement part. It has
an entropy-based similarity measure in construction and refinement of
the concept hierarchy. It makes use of an attribute-based metric to
obtain concept stability, and an instance-based metric for tree
stability.

The system has been implemented, and experiments have been performed in
several domains to measure the applicability of the system. The results
showed that the conceptual validity of the concept hierarchy is high,
and the effect of order of instance arrivals is minimized.
---------------------------------------------------------------------

Keywords:
   Inductive learning, Conceptual clustering, Incremental learning.