Ren, Yongli, Ye, Yangdong and Li, Gang 2008, The density connectivity information bottleneck, in Proceedings of the 9th International Conference for Young Computer Scientists, IEEE Computer Society, Piscataway, N.J., pp. 1783-1788.
(Some files may be inaccessible until you login with your Deakin Research Online credentials)
Clustering with the agglomerative Information Bottleneck (aIB) algorithm suffers from the sub-optimality problem, which cannot guarantee to preserve as much relative information as possible. To handle this problem, we introduce a density connectivity chain, by which we consider not only the information between two data elements, but also the information among the neighbors of a data element. Based on this idea, we propose DCIB, a Density Connectivity Information Bottleneck algorithm that applies the Information Bottleneck method to quantify the relative information during the clustering procedure. As a hierarchical algorithm, the DCIB algorithm produces a pruned clustering tree-structure and gets clustering results in different sizes in a single execution. The experiment results in the documentation clustering indicate that the DCIB algorithm can preserve more relative information and achieve higher precision than the aIB algorithm.