150 likes | 296 Views
Iterative shrinking method for generating clustering. Olli Virmajoki , Pasi Fränti and Timo Kaukoranta. UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE FINLAND. TURKU CENTRE FOR COMPUTER SCIENCE UNIVERSITY OF TURKU FINLAND. Problem setup.
E N D
Iterative shrinking method for generating clustering Olli Virmajoki, Pasi Fränti andTimo Kaukoranta UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE FINLAND TURKU CENTRE FOR COMPUTER SCIENCE UNIVERSITY OF TURKU FINLAND
Problem setup Given N data vectors X={x1, x2, …, xN}, partition the data set into M clusters 1. Goal of clustering: find the location of the clusters. 2. Goal of VQ: approximate the original data by the codebook.
Algorithms for solving the clustering PNN: Pairwise Nearest Neigbor method • Existing method • Merges clusters IS: Iterative shrinking method • New method • Removes clusters
Local optimization of the PNN Merge cost: Find minimum cost pair to be merged:
Local optimization of the IS Finding secondary cluster: Removal cost of single vector:
Different update strategies Time complexities: O(N2) .. O(N2 log2N)
Comparison of run times (Extensive update)
Number of distance calculations Approximately the same >2 times slower No much difference
Comparison of MSE-values • Additional result (M=256): • Genetic Algorithm with IS as crossover: 160.77
Conclusions • Slower but better clustering algorithm. • Local optimization applied in every step. • Preliminary results: BEST known clustering algorithm (in minimizing MSE) when used as crossover in Genetic Algorithm !!!