A novel rough semi-supervised k-means algorithm for text clustering

Since many attribute values of high-dimensional sparse data are zero, we combine the approximation set of the rough set theory with the semi-supervised k-means algorithm to propose a rough set-based semi-supervised k-means (RSKmeans) algorithm. Firstly, the proportion of non-zero values is calculated by a few labelled data samples, and a small number of important attributes in each cluster are selected to calculate the clustering centres. Secondly, the approximation set is used to calculate the information gain of each attribute. Thirdly, different attribute values are partitioned into the corresponding approximate sets according to the comparison of information gain with the upper approximation and boundary threshold. Then, the new attributes are increased and the above process is continued to update the clustering centres. The experimental results on text data show that the RSKmeans algorithm can help find the important attributes, filter the invalid information, and improve the performances significantly.