Clustering Algorithm Based on the Distributions of Intrinsic Clusters
一种基于语料特性的聚类算法

In finding a flexible approach to solve the model misfit problem, a clustering algorithm based on the distributions of intrinsic clusters (CADIC) is proposed, which implicitly integrates distribution characteristics into the clustering framework by applying rescaling operations. In the clustering process, a set of discriminative directions are chosen to construct the CADIC coordinate, under which the distribution characteristics are analyzed in order to design rescaling functions. Along every axis, rescaling functions are applied to implicitly normalize the data distribution such that more reasonable clustering decisions can be made. As a result, the reliability of clustering decisions is improved. The time complexity of CADIC remains the same as K-means by using a K-means-like iteration strategy. Experiments on well-known benchmark evaluation datasets show that the framework of CADIC is reasonable, and its performance in text clustering is comparable to that of state-of-the-art algorithms.