In the original IB (Information Bottleneck) algorithms

all the data points are employed to learn the cluster patterns.However

in many real-world applications

some data show clear coherent behavior and can be summarized well

while some data present weak tendencies to be assigned to any particular pattern.For such situations

this paper proposes a DSIB (Data Selection Information Bottleneck) algorithm which has the ability to select data points with clear coherent behavior and find their corresponding cluster patterns.To realize this goal

the DSIB algorithm takes the information loss as the data selection criterion

which is generated when we try to compress the data point into one of the clusters.The DSIB algorithm adopts sequential draw-and-merge procedure to select the data and learn the cluster patterns.This learning process can take full account of each datum's natural pattern.Experimental results show that with the improvement of the data selection criterion

the DSIB algorithm can improve the clustering precision while the expense of the recall is small.In our evaluation

the DSIB algorithm is found to be consistently superior to all the other clustering methods we examine.