Abstract:

 A novel rough K-means clustering algorithm was presented  based on the weight of exemplar density to overcome the drawback of selecting initial seeds randomly of available rough K-means algorithms. A new density function was defined for each sample according to the denseness of samples around it without any arbitrary parameter, and the top K samples with higher density and far away from each other were selected as initial centers of rough K-means clustering algorithm. Further more the new weight was defined for each exemplar according to the value of the new density function, so that the better could croids of each cluster could be calculated out without influenced by noisy data. Experiments on six UCI data sets and on synthetically geterated  data sets  with noise points proved that our algorithm got a better clustering result, and had a strong anti-interference performance for noise data.
 

Key words: clustering algorithm; rough K-means; clustering center; weight; density