(Georgia Institute of Technology, 1999)
Omiecinski, Edward; Navathe, Shamkant B.; Ezquerra, Norberto F.; Ordońẽz, Carlos
This paper presents a clustering algorithm to discover low and high density
regions in subspaces of multidimensional data for Data Mining applications.
High density regions generally refer to typical cases, whereas low density
regions indicate infrequent and thus rare cases. For typical applications
there is a large number of low density regions and a few of these are
interesting. Regions are considered interesting when they have a minimum
"volume" and involve some maximum number of dimensions. Our algorithm
discovers high density regions (clusters) and low density regions (outliers,
negative clusters, holes, empty regions) at the same time. In particular,
our algorithm can find empty regions; that is, regions having no data
points. The proposed algorithm is fast and simple. There is a large variety
of applications in medicine, marketing, astronomy, finance, etc, where
interesting and exceptional cases correspond to the low and high density
regions discovered by our algorithm.
(Georgia Institute of Technology, 1999)
Omiecinski, Edward; Navathe, Shamkant B.; Ammar, Mostafa H.; Donahoo, Michael J.; Malik, Sanjoy; Yee, Wai Gen
Replication of data on portable computers is a new DBMS technology aimed at
catering to a growing population of mobile database users. Clients can
download data items such as email, or sales data from a server onto these
machines, per use it during commutes, and return any modifications to the
server at the end of the day. In this paper, we describe how the servers in
these systems generally process update information for clients and reveal a
scalability problem--server processing increases quadratically with respect
to increasing numbers of clients. We develop a cost model, and propose a
solution based on heuristics. By aggregating client interests into
datagroups, based on notions such as interest overlap, we can reduce server
cost. These techniques are attractive because they are simple and
computationally cheap. Simulations show that even simple techniques may
yield significant performance improvements.