Naive Bayes clusterer

摘要

It is common practice to process labeled data with classifiers and unlabeled ones with clusterers. The idea of combining supervised and unsupervised learning methods to process data in a more efficient way leads to semi-supervised learning methods, which effectively utilizes both labeled and unlabeled data. However, the thought of processing unlabeled data with a combination of supervised and unsupervised methods to achieve better efficiency and accuracy calls for the proposal of a brand new clustering model that we call Naive Bayes Clusterer Model (NBCM). In this paper, we built NBCM based on an optimization problem of conditional probability and solved it by partitioning the hyperedges consists of learnt label groups from unsupervised learning. Extensive experiments were conducted and results on real data show that NBCM outperforms all other base clustering algorithms compared with the highest level of accuracy.