In the field of pattern recognition, the traditional supervised learning methods and unsupervised learning methods are not always suitable for the practical applications. In some applications, the data obtained is neither no-information-given nor all-information-given. In addition, the data obtained usually contains some noises due to many interference factors in practical collection procedure and these noises are of great influence on the traditional clustering methods. In order to overcome the two problems mentioned above, based on the classical Maximal Entropy Clustering (MEC), we propose a semi-supervised MEC algorithm based on the maximized central distance and the compensation term for membership, i.e., CMsSMEC algorithm. The experimental results on benchmarking UCI data sets show that it has a better performance than the traditional unsupervised clustering method.