期刊
KNOWLEDGE-BASED SYSTEMS
卷 135, 期 -, 页码 147-158出版社
ELSEVIER
DOI: 10.1016/j.knosys.2017.08.010
关键词
Document clustering; Concept factorization; Manifold regularization
资金
- Science and Technology Development Fund (FDCT) of Macao SAR [128/2013/A, 124/2014/A3]
- National Natural Science Foundation of China [61602540, 61722304]
- Guangdong Natural Science Funds for Distinguished Young Scholar [2014A030306037]
- Pearl River S&T Nova Program of Guangzhou [201610010196]
- Guangdong Program for Support of Top-notch Young Professionals [2014TQ01X144]
Document clustering is an important tool for text mining with its goal in grouping similar documents into a single cluster. As typical clustering methods, Concept Factorization (CF) and its variants have gained attention in recent studies. To improve the clustering performance, most of the CF methods use additional supervisory information to guide the clustering process. When the amount of supervisory information is scarce, the improved performance of CF methods will be limited. To overcome this limitation, this paper proposes a novel regularized concept factorization (RCF) algorithm with dual connected constraints, which focuses on whether two documents belong to the same class (must-connected constraint) or different classes (cannot-connected constraint). RCF propagates the limited constraint information from constrained samples to unconstrained samples, allowing the collection of constraint information from the entire data set. This information is used to construct a new data similarity matrix that concentrates on the local discriminative structure of data. The similarity matrix is incorporated as a regularization term in the CF objective function. By doing so, RCF is able to make full use of the supervisory information to preserve the local structure of the data set. Thus, the clustering performance will be improved significantly. Our experiments on standard document databases demonstrate the effectiveness of the proposed method. (C) 2017 Elsevier B.V. All rights reserved.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据