一种结合主动学习的半监督文档聚类算法

Efficiently Active Learning for Semi-Supervised Document Clustering

半监督文档聚类,即利用少量具有监督信息的数据来辅助无监督文档聚类,近几年来逐渐成为机器学习和数据挖掘领域研究的热点问题.由于获取大量监督信息费时费力,因此,国内外学者考虑如何获得少量但对聚类性能提高显著的监督信息.提出一种结合主动学习的半监督文档聚类算法,通过引入成对约束信息指导DBSCAN的聚类过程来提高聚类性能,得到一种半监督文档聚类算法Cons-DBSCAN.通过对约束集中所含信息量的衡量和对DBSCAN算法本身的分析,提出了一种启发式的主动学习算法,能够选取含信息量大的成对约束集,从而能够更高效地辅助半监督文档聚类.实验结果表明,所提出的算法能够高效地进行文档聚类.通过主动学习算法获得的成对约束集,能够显著地提高聚类性能.并且,算法的性能优于两个代表性的结合主动学习的半监督聚类算法.

Semi-Supervised document clustering and employing limited prior knowledge to aid in unsupervised clustering, have recently become a topic of significant interest to data mining and machine learning communities. Because receiving supervised data may be expensive, it is important to attain the most informative knowledge to improve the clustering performance. This paper presents a semi-supervised document clustering algorithm with active learning for pairwise constraints, aiming at getting improved clustering performance. The semi-supervised document clustering algorithm is a constrained DBSCAN （cons-DBSCAN） algorithm, which incorporates pairwise constraints to guide the clustering process in DBSCAN. Basing on measure of constraint set utility and analysis of DBSCAN algorithm, an active learning approach is proposed to select informative document pairs for obtaining user feedbacks. Experimental results show that this proposed approach is effective in document clustering. The clustering performance of active Cons-DBSCAN has dramatically improved with selected pairwise constraints. Moreover, the proposed approach performs better than the two representative methods.