Enhanced Mode Clustering


Yen-Chi Chen

11:00:00 - 13:00:00

103 , Mathematics Research Center Building (ori. New Math. Bldg.)

Mode clustering is a nonparametric method for clustering which uses the basins of attraction of the modes of a density estimator to define clusters. There are several advantages to mode clustering over other commonly used methods. First, there is a clear population quantity being estimated, namely, the Morse complex of the true density. Second, the computation can be easily done by the mean shift algorithm. Third, there is a single tuning parameter to choose: the bandwidth of the density estimator. Fourth, it is theoretically well understood since it depends only on density estimation and mode estimation

We provide several enhancements to mode clustering:
(i) we create soft versions of mode clustering,
(ii) we introduce a method for measuring connectivity between mode clustering,
(iii) we propose a method for choosing the bandwidth,
(iv) we propose a method for merging overlapping clusters and
(v) we introduce a method for visualizing the clusters.