Co-Clustering-Based Clustering and Segmentation for Pattern Discovery from Time Course Data

Abstract

Time course data may inherit critical temporal 
ordering in contiguous (i.e., neighboring) time slot. Traditional 
one-way k-means clustering algorithms handle time points 
independently, ignoring the internal time locality. Although 
co-clustering algorithms can discover latent local patterns, the 
discovered patterns are not necessary to be in a continuous time 
order. Therefore, this paper targets to extend an existing 
co-clustering framework to be applicable to time course data so 
that time-dependent local segment patterns over specific 
intervals can be captured. While following the general 
co-clustering framework of the alternating optimization process, 
the proposed algorithms employ clustering on instance 
dimension and segmentation on time dimension. Both batch and 
incremental updates at boundary time points are proposed to 
search for a sequence of time segments. Eight time course 
datasets and two specific data normalization schemes are 
considered in the experimental study.  Clustering similarity 
performance among k-means, one existing co-clustering, and 
the two proposed clustering segmentation algorithms is 
compared.