Multi-label feature selection via feature manifold learning and sparsity regularization
- PDF / 1,966,306 Bytes
- 14 Pages / 595.276 x 790.866 pts Page_size
- 95 Downloads / 188 Views
ORIGINAL ARTICLE
Multi‑label feature selection via feature manifold learning and sparsity regularization Zhiling Cai1,2 · William Zhu1
Received: 7 July 2016 / Accepted: 20 January 2017 © Springer-Verlag Berlin Heidelberg 2017
Abstract Multi-label learning deals with data associated with different labels simultaneously. Like traditional singlelabel learning, multi-label learning suffers from the curse of dimensionality as well. Feature selection is an efficient technique to improve learning efficiency with high-dimensional data. With the least square regression model, we incorporate feature manifold learning and sparse regularization into a joint framework for multi-label feature selection problems. The graph regularization is used to explore the feature geometric structure for gaining a better regression coefficient matrix which reflects the importance of varying features. Besides, the 𝓁2,1-norm is imposed on the sparsity term to guarantee the sparsity of the regression coefficients. Furthermore, we design an iterative updating algorithm with proved convergence to tackle the aforementioned formulated problem. The proposed method is validated in six publicly available data sets from real-world applications. Finally, extensively experimental results demonstrate its superiority over the compared state-of-the-art multi-label feature selection methods. Keywords Multi-label learning · Feature selection · Supervised learning · Graph regularization · 𝓁2,1-norm
* William Zhu [email protected] Zhiling Cai [email protected] 1
Institute of Fundamental and Frontier Sciences, University of Electronic Science and Technology of China, Chengdu, China
2
Lab of Granular Computing, Minnan Normal University, Zhangzhou, China
1 Introduction In traditional single-label learning, each instance is associated with one label to indicate its concept class belongingness. However, in many real-world applications, one object usually attaches more than one label implying that an instance corresponds to a set of labels. For example, in text categorization [34, 41], a document may belong to multiple topics, such as government and health. In image annotation [3, 67], an image may have several semantic classes, such as beach and urban. In functional genomics [15, 65], each gene may match a set of functional classes including matabolism, transcription and protein synthesis. From the above consideration, the paradigm of multi-label learning [24, 58, 61, 64] naturally emerges with increasing attention. In learning task, high dimensionality [42, 54–56] often causes serious negative problems, such as computational burden, poor performance and over-fitting. To ease these problems, dimensionality reduction methods have been developed [6, 19, 28, 48]. The two main approaches to dimensionality reduction are feature extraction and feature selection. The former aims to map the original features into a low-dimensional subspace via a certain transformation [25, 60, 68], while the latter is to select a small feature subset given a ce
Data Loading...
