Abstract: Semi supervised clustering algorithms have recently received a significant amount of attention in the machine learning and data mining of communities. All of current algorithms use attribute-value languages to represent knowledge. Attributcvaluc languages have inherent drawback for represent complex structured data. However, knowledge representation language Escher based on higher-order logic, can represent complex structured data. With Escher as knowledge representation formalism, firstly, when prior knowledge is pairwise constraints between instances, the method of initializing the K-Means algorithm was proposed; secondly, when the number r of cluster centers which can be initialized with incomplete knowledge is less than K, the algorithms MSS-KMeans and SMSS-KMeans were proposed to initialize the rest K一r cluster centers. Finally, the empirical study carried out on datasets of complex structure data showed the feasibility of the presented algorithms. I}he final experimental results demonstrate the comparability between the presented algorithms and the known algorithms based on attribute-value language.

Key words: Higher-order logic, Template, Semi-supervised clustering, Complex structure data, Prior knowledge