Cross-regression for multi-view feature extraction
Introduction
With the increase of information collection channels, data is often collected from diverse feature extractors. For example, an image can be represented by different types of papers, and a web page can be represented by hyperlinks and content texts. These different types of samples are called multi-view data. Obviously, multi-view data may characterize different specific information and contains more information than the single-view data. Multi-view learning (MVL) [1], [2], [3], [4], [5], [6] is proposed to integrate compatible and complementary information among different views, which has better performance than traditional single-view learning. Recently, MVL has been widely expanded to many fields, such as multi-view multi-instance learning [7], multi-view clustering [8], [9] and multi-view feature extraction, etc. [10].
In practical applications, the samples usually locate in a high dimensional space, such as image classification [11] or clustering [12]. If we deal with the high-dimensional samples directly, it will lead to the “curse of dimensionality” [13]. So, feature extraction is always a crucial processing step to obtain a tractable low-dimensional representation [14], [15], [16]. Traditional feature extraction methods include Principal Component Analysis (PCA) [17] and Linear Discriminant Analysis (LDA) [18]. However, they are only suitable for single-view data and not for multi-view data. For multi-view data, various multi-view feature extraction (MvFE) methods are proposed, which exploit the correlation information from multiple views when extracting features. Canonical correlation analysis (CCA) [19] and Partial Least Squares (PLS) [20] are two typical methods, which aim at maximizing between-view correlation and covariance respectively. Sun and Wang et al. proposed locality preserving canonical correlation analysis (LPCCA) [21] and a new LPCCA (ALPCCA) [22] respectively. LPCCA and ALPCCA find the low-dimensional embedding by preserving local neighbor information. The main difference is that the former maximizes local canonical correlation coefficient, while the later explores extra cross-view correlations between neighbors. Nevertheless, the number of neighbors in both methods is manually chosen by experience, which affects the final results. Thus, Zu et al. proposed canonical sparse cross-view correlation analysis (CSCCA) [23] which combines sparse reconstruction and LPCCA to explore local intrinsic geometric structure automatically. Zhu et al. pointed out that CSCCA neglects the weights of data, as the difference among samples is not well modeled. Therefore, a weight-based CSCCA (WCSCCA) [24] was proposed. WCSCCA measures the correlation between two views using the weights of data and the cross-view information. For two-view feature extraction, Zhao et al. proposed co-training locality preserving projections (Co-LPP) [25] which aims at finding a low-dimensional embedding such that the local neighbor structures of two views are maximumly compatible.
Note that, all of the above methods are only suitable for two-view scenario. In order to deal with multi-view scenario, Foster et al. proposed Multi-view CCA (MCCA) [26] which tries to find a common space by maximizing the total canonical correlation coefficients between any two views. As a further extension, Cao et al. proposed multi-view PLS (MvPLS) [27]. By unifying Laplacian eigenmaps (LE) [28] and multi-view learning, Xia et al. proposed multi-view spectral embedding (MSE) [29]. MSE finds a subspace in which the low-dimensional embedding is sufficiently smooth. Wang et al. adopted a novel locality linear embedding scheme to develop a new method named multi-view reconstructive preserving embedding (MRPE) [30]. Combining sparse reconstruction and co-regularized scheme, a co-regularized multi-view sparse reconstruction embedding (CMSRE) [31] was proposed. A common characteristic of MSE, MRPE and CMSRE is that they find the low-dimensional embedding directly, and it is unclear how to calculate the low-dimensional presentation of a new point. Therefore, sparsity preserving multiple canonical correlation analysis (SPMCCA) [32] and graph multi-view canonical correlation analysis (GMCCA) [33] were proposed.
Most MvFE methods have two characteristics: (1). They mainly explore the cross-view correlations in the projected subspace without considering the correlation information in original high-dimensional space; (2). They are sensitive to the outliers due to using L2-norm or F-norm [34], [35], [36]. Recently, ridge regression (RR) has made a breakthrough, which uses the previous information (original data or label information) directly in the regression models. There appears many feature extraction methods based on RR for single-view data, such as robust discriminant regression (RDR) [37] and generalized robust regression (GRR) [38]. They aim at finding a robust subspace to maintain the local manifold structure.
Inspired by the regression strategy for feature extraction, this paper constructs a novel cross-regression regularization term to discover the relationship between multiple views in original high-dimensional space. Firstly, minimizing the proposed regularization term aims at seeking a set of projection matrices to transform the samples from different views into a common low-dimensional subspace. Then, another set of projection matrices is introduced to transform the low-dimensional samples back to the original high-dimensional space. Minimizing our proposed cross-regression regularization term is to minimize the distance between original data and the projected high-dimensional data. Finally, two novel MvFE methods, cross-regression for multi-view feature extraction (CRMvFE) and robust cross-regression for multi-view feature extraction (RCRMvFE) are proposed. The main contributions of this paper can be concluded as follows:
(1) A novel cross-regression regularization term is designed and a regression method with this regularization term named CRMvFE is proposed. CRMvFE makes better use of previous information (high-dimensional data) than traditional MvFE methods, and it explores the correlations from multiple views and single-view.
(2) A robust CRMvFE (RCRMvFE) is proposed for extracting robust subspace embedding by adding L2,1-norm instead of F-norm. Furthermore, the influence of outliers can be effectively reduced by using L2,1-norm.
(3) An effective iterative algorithm for solving RCRMvFE is proposed. Theoretical analysis about convergence of the algorithm and the relationship between RCRMvFE and CRMvFE are discussed.
(4) Experimental results on image datasets and hyperspectral image datasets demonstrate the validity and advantage of CRMvFE and RCRMvFE.
The rest of the paper is organized as follows: In Section 2, some related works are introduced. Section 3 presents the proposed CRMvFE. An improved RCRMvFE and its theoretical analysis are presented in Section 4. The experimental results and the conclusion are given in Section 5 and Section 6 respectively.
Section snippets
Related works
In this section, we briefly introduce two typical multi-view feature extraction methods: multi-view canonical correlationanalysis (MCCA) and graph multi-view canonical correlation analysis (GMCCA).
Given the data matrices: where represents the sample matrix from the th view, represents the th sample of the th view, is the feature dimension in th view. Given , multi-view feature extraction aims to find the projection matrices
Model of CRMvFE
A new MvFE named as cross-regression for multi-view feature extraction (CRMvFE) is proposed in this section. CRMvFE preserves the correlation between multiple views directly by introducing a novel cross-regression regularization term and explores the correlation in single-view itself. For the data matrices given by (1), the sketch of CRMvFE is illustrated in Fig. 1.
CRMvFE finds the projection matrices , by solving the following optimization problem:
Robust Cross-regression for Multi-view Feature Extraction (RCRMvFE)
CRMvFE proposed in the above section utilizes F-norm as the metric, which is sensitive to outliers or noises. Therefore, a novel robust CRMvFE (RCRMvFE) is constructed, where L2,1-norm is utilized.
Experiments
In this section, three image datasets and two hyperspectral image datasets are applied to validate the effectiveness of CRMvFE and RCRMvFE. We compare our methods with five well known multi-view feature extraction methods: CoLPP, MCCA, SPMCCA, MvPLS and GMCCA. All the experiments are implemented in Matlab 2016a on a PC with 16RAM and the code of the proposed methods can be downloaded from http://www.scholat.com/zhangjinxin.
Conclusion
In this paper, we propose two novel methods called CRMvFE and RCRMvFE for MvFE. A novel cross-regression regularization term is designed to explore the correlations among multiple views and simultaneously obtain the low-dimensional projection matrices for each view. In addition, in order to reduce the influence of outliers or noises, we propose RCRMvFE which combines CRMvFE and L2,1-norm. Theoretical analysis about convergence of the algorithm and the relationship between CRMvFE and RCRMvFE are
CRediT authorship contribution statement
Jinxin Zhang: Data curation, Formal analysis, Investigation, Software, Methodology, Validation, Visualization, Writing - original draft, Writing - review & editing. Ling Jing: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing - original draft, Writing - review & editing. Junyan Tan: Conceptualization, Data curation, Formal analysis, Investigation, Software, Methodology, Validation,
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work is supported by the National Natural Science Foundation of China (Nos. 11671032)
References (38)
- et al.
Incremental multi-view spectral clustering
Knowl.-Based Syst.
(2019) - et al.
One-step kernel multi-view subspace clustering
Knowl.-Based Syst.
(2019) - et al.
A multitask multiview clustering algorithm in heterogeneous situations based on LLE and LE
Knowl. Based Syst.
(2019) - et al.
A study of graph based system for multiview clustering
Knowl. Based Syst.
(2019) - et al.
2DRLPP Robust two dimensional locality preserving projection with regularization
Knowl.-Based Syst.
(2019) - et al.
Low rank local tangent space embedding for subspace clustering
Inform. Sci.
(2020) - et al.
An efficient selector for multigranularity attribute reduction
Inform. Sci.
(2019) - et al.
Interactive dimensionality reduction using similarity projections
Knowl.-Based Syst.
(2019) - et al.
Interactive dimensionality reduction using similarity projections
Knowl.-Based Syst.
(2018) - et al.
Locality preserving CCA with applications to data visualization and pose estimation
Image Vis. Comput.
(2007)
Multi-view reconstructive preserving embedding for dimension reduction
Soft Comput.
Co-regularized multi-view sparse reconstruction embedding for dimension reduction
Neurocomputing
A joint-L2,1-norm-constraint-based semi-supervised feature extraction for RNA-Seq data analysis
Neurocomputing
Multi-graph fusion for multi-view spectral clustering
Knowl.-Based Syst.
Multi-view locality low-rank embedding for dimension reduction
Knowl.-Based Syst.
Efficient robust model fitting for multistructure data using global greedy search
IEEE Trans. Cybern.
Collaborative weighted multi-view feature extraction
Eng. Appl. Artif. Intell.
Clustering based multiple instance learning with multiview feature
Expert Syst. Appl.
Weighted low-rank representation-based dimension reduction for hyperspectral image classification
IEEE Geosci. Remote Sens. Lett.
Cited by (15)
Relaxed multi-view discriminant analysis
2024, Engineering Applications of Artificial IntelligenceMulti-view robust regression for feature extraction
2024, Pattern RecognitionMulti-view clustering via optimal transport algorithm
2023, Knowledge-Based SystemsGeneralized multiview regression for feature extraction
2023, Information SciencesMultiview Jointly Sparse Discriminant Common Subspace Learning
2023, Pattern RecognitionOrthogonal multi-view analysis by successive approximations via eigenvectors
2022, NeurocomputingCitation Excerpt :To take full advantage of multi-view data, multi-view learning has attracted increasing attention due to its wide applications such as dimensionality reduction [1], cross-view recognition [2,3], clustering [4,5], classification [6], and multi-label learning [7,8]. Many learning criteria have been explored to capture the relations among multiple views including subspace learning methods [9,10], tensor approaches [11,12] and the deep learning [13–15]. Although great progress has been made by existing multi-view learning methods, there are still challenges.