基于不完备标签数据的半监督聚类算法

Semi-Supervised Clustering Algorithm Based on Incomplete Labeled Data

针对seeded-K-means和constrained-K-means算法要求标签数据类别完备的限制,本文提出了基于不完备标签数据的半监督K-means聚类算法,重点讨论了未标签类别初始聚类中心的选取问题。首先给出了未标签类别聚类中心最优候选集的定义,然后提出了一种新的未标签类别初始聚类中心选取方法,即采用K-means算法从最优候选集中选取初始聚类中心,最后给出了基于新方法的半监督聚类算法的完整描述,并通过实验测试对新算法的有效性进行了验证。实验结果表明本文所提算法在执行速度和聚类效果上都优于现有算法。

For the seeded-K-means and constrained-K-means algorithm limitations that complete category information in labeled data is required, this paper put forword an semi-supervised K-means clustering algorithm based on incomplete labeled data, focused on selection of the initial cluster center of unlabeled category. We gave a definition of the Best Candidate Set of cluster center of unlabeled category, proposed a new method that selecting initial cluster center of unlabeled category from the Best Candidate Set using K-means. Finally, a complete description of semi-supervised clustering algorithm based on the new method is given, the validity of the new algorithm is verified by experiment. Experimental results show that the proposed algorithm is superior to existing algorithms not only in clustering effect and in execution speed.