The PreDeCon clustering algorithm finds arbitrarily shaped clusters in high-dimensional feature spaces, which remains an active research topic with many potential applications. However, it suffers from poor runtime performance, as well as a lack of user interaction. Our new method AnyPDC introduces a novel approach to cope with these problems by casting PreDeCon into an anytime algorithm. In this anytime scheme, it quickly produces an approximate result and iteratively refines it toward the result of PreDeCon at the end. AnyPDC not only significantly speeds up PreDeCon clustering but also allows users to interact with the algorithm during its execution. Moreover, by maintaining an underlying cluster structure consisting of so-called primitive clusters and by block processing of neighborhood queries, AnyPDC can be efficiently executed in parallel on shared memory architectures such as multi-core processors. Experiments on large real world datasets show that AnyPDC achieves high quality approximate results early on, leading to orders of magnitude speedup compared to PreDeCon. Moreover, while anytime techniques are usually slower than batch ones, the algorithmic solution in AnyPDC is actually faster than PreDeCon even if run to the end. AnyPDC also scales well with the number of threads on multi-cores CPUs.
|Title of host publication||Transactions on Large-Scale Data- and Knowledge-Centered Systems XXXIX|
|Publication status||Published - 23 Nov 2018|
|Name||Transactions on Large-Scale Data- and Knowledge-Centered Systems|