Elsevier

Information Fusion

Volume 41, May 2018, Pages 119-128
Information Fusion

Multiview dimension reduction via Hessian multiset canonical correlations

https://doi.org/10.1016/j.inffus.2017.09.001Get rights and content

Highlights

  • We integrates Hessian regularization into multiset CCA for multiview dimension reduction.

  • We provide the optimization of HesMCC algorithm.

  • We conduct extensive experiments to verify the proposed HesMCC and the results are promising.

Abstract

Canonical correlation analysis (CCA) is a main technique of linear subspace approach for two-view dimension reduction by finding basis vectors with maximum correlation between the pair of variables. The shortcoming of the traditional CCA lies that it only handles data represented by two-view features and cannot reveal the nonlinear correlation relationship. In recent years, many variant algorithms have been developed to extend the capability of CCA such as discriminative CCA, sparse CCA, kernel CCA, locality preserving CCA and multiset canonical correlation analysis (MCCA). One representative work is Laplacian multiset canonical correlations (LapMCC) that employs graph Laplacian to exploit the nonlinear correlation information for multiview high-dimensional data. However, it possibly leads to poor extrapolating power because Laplacian regularization biases the solution towards a constant function. In this paper, we present Hessian multiset canonical correlations (HesMCC) for multiview dimension reduction. Hessian can properly exploit the intrinsic local geometry of the data manifold in contrast to Laplacian. HesMCC takes the advantage of Hessian and provides superior extrapolating capability and finally leverage the performance. Extensive experiments on several popular datasets for handwritten digits classification, face classification and object classification validate the effectiveness of the proposed HesMCC algorithm by comparing it with baseline algorithms including TCCA, KMUDA, MCCA and LapMCC.

Introduction

Many computer vision tasks employ multiple information rather than a single representation to achieve more robust performance. For example, an image can be described with color, shape, and texture features [1], [2], [3]. The extracted visual features usually have high dimensions of up to hundreds or thousands, which often causes the problem called the curse-of-dimensionality [4], [5]. Hence multiview dimension reduction algorithms [6], [7] have been subsequently proposed with the purpose of finding an appropriate low-dimensional feature subspace from multiview high-dimensional features. Canonical correlation analysis (CCA) [8] is one of the most representative techniques and has been widely applied to many multiview learning applications including classification, retrieval, regression and clustering.

Canonical correlation analysis (CCA) proposed by Hotelling [8] seeks a pair of linear transformation for two view high-dimensional features such that the corresponding low-dimensional projections are maximally correlated. In recent years, dozens of CCA extensions have been developed and these algorithms can be roughly categorised into the following four groups namely discriminative CCA, sparse CCA, kernel CCA and locality preserving CCA.

Discriminative CCA [9], [10], [11] considers the combination of within-class/between-class information and correlated information of training samples to improve the discrimination of the low-dimensional subspace. For example, Kim et al. [9] developed a linear discriminant function of CCA similarly to linear discriminant analysis (LDA) by maximizing the canonical correlations of within-class sets and minimizing the canonical correlations of between-class sets. Wang et al. [10] proposed an unsupervised discriminant CCA based on spectral clustering by utilizing the correlation information between the samples in the same class including the correlation between paired data, the correlation across views, and the correlation within views. Sun et al. [11] added the uncorrelation constraint into multiview linear discriminant analysis (MLDA) and proposed the multiview ULDA (MULDA) to seek the project directions with minimum redundancy. Besides, they also extended the MULDA to its kernel version, i.e., kernel multiview uncorrelated discriminant analysis (KMUDA), to tackle the nonlinear case.

Sparse CCA [12], [13], [14] seeks two sparse canonical vectors to maximize the correlation between the two views. Hardoon and Shawe-Taylor [12] proposed a sparse CCA by minimizing the number of features while maximizing the correlation between the primal view and the dual view. Witten et al. [13] employed penalized matrix decomposition to optimize l1-regularized sparse CCA. And Chen et al. [14] developed a structured sparse CCA by incorporating the structured-sparsity-inducing penalty to exploit both pre-given group structure and unknown group structure.

Kernel CCA [15], [16], [17], [18] finds maximally nonlinear correlations via an implicit nonlinear mapping to increase the flexibility of the feature selection. Lai and Fyfe [15] used a kernel trick to map the data to a high-dimensional feature space and then performed the traditional CCA on the feature space. Akaho [17] developed regularized kernel CCA to avoid the overfitting by incorporating regularization technique into kernel CCA. Zhu et al. [18] utilized a mixture of kernels rather than a single kernel as the nonlinear mapping and improved both interpolation ability and extrapolation ability of kernel CCA.

Locality preserving CCA incorporates locality information into CCA and then preserves the local manifold structure [19] while obtaining the canonical correlation [20]. Sun and Chen [21] decomposed the global nonlinear structure into many locally linear ones, then conducted linear CCA on each small neighborhood and finally integrated the local sub-problems to obtain the canonical correlation.

Although CCA and its variants have achieved practical success, they are in nature a two-view dimension reduction techniques and cannot handle the problem of multiview dimension reduction that often exists in many real-world applications. Multiset canonical correlation analysis (MCCA) [22] then has been developed to tackle the correlation analysis of multiple variables. Luo et al. [23] proposed tensor CCA to maximize the correlation between the multiple canonical variables by finding the best rank-1 approximation of the high-order covariance tensor. And recently, Yuan et al. [24] proposed Laplacian multiset canonical correlations (LapMCC) to discover the nonlinear correlations among multiview features by combining many locally within-view and between-view correlations together.

In spite of LapMCC has achieved promising performance compared with the traditional MCCA algorithms, it has been identified that Laplacian will bias the solution towards a constant function and then lead to poor extrapolating power [25]. On the other hand, Hessian predicts a high order derivatives and has a richer null space, which makes it can steer the prediction varying smoothly along the underlying manifold. In this paper, we integrate Hessian into the multiset canonical correlations and derive Hessian multiset canonical correlations (HesMCC). HesMCC takes the advantage of Hessian and provides superior extrapolating capability. Therefore, HesMCC can significantly leverage the performance. Fig. 1 graphically demonstrates the whole procedure of the proposed HesMCC. Briefly speaking, the contribution of this paper includes the following three folds: (1) we derive the locality preserving canonical correlation by using Hessian; (2) we further formulate HesMCC to analyze multiset canonical correlation; (3) we present the algorithm of HesMCC and conduct extensive experiments to verify the proposed HesMCC.

Finally, we carefully implement HesMCC for multiview dimension reduction and conduct extensive experiments on USPS database for handwritten digits classification, Yale-B database and ChokePoint database for face recognition and ETH-80 database for object classification respectively. We also compare HesMCC with TCCA, KMUDA, MCCA and LapMCC algorithms to evaluate the performance of HesMCC. The experimental results verify the effectiveness of HesMCC by comparison with the baseline algorithms.

The rest of this paper is organized as follows. We firstly review some related works in Section 2. Then we derive the proposed HesMCC algorithm in Section 3. Section 4 details the implementation of HesMCC. And experimental results are discussed in Section 5, followed by the conclusion in Section 6.

Section snippets

Related work

In this section, we first briefly summarize the related work for the multiview canonical correlation analysis including multiset canonical correlation analysis (MCCA), multiset integrated canonical correlation analysis (MICCA), tensor canonical correlation analysis (TCCA), and Laplacian multiset canonical correlation (LapMCC). And then, we give a brief null space description of Laplacian and Hessian.

Hessian multiset canonical correlations (HesMCC)

In this section, we firstly describe the local preserve CCA by Hessian and then extend it to HesMCC for the multiview case.

Algorithm

In this section, we describe the optimization of problem (4) using Lagrange multiplier technique. Consider λ as the Lagrange multiplier, the Lagrangian function of problem (4) can be written asF(α¯,λ)=i=1mj=1mαiTSijHαjλ(i=1mαiTSiiHαi1)

The partial derivative of F(α¯,λ)w.r.t. αi has the following expressionFαi=2j=1mSijHαi2λSiiHαi,i=1,2,,m.

Set Fαi=0, we havej=1mSijHαi=λSiiHαi,i=1,2,,m.

The m equations of (7) can be finally reformulated asSHα˜=λSDHα˜where SH is a block matrix with

Experiments

To evaluate the performance of HesMCC, we implement HesMCC for multiview dimension reduction and compare it with the traditional MCCA and the representative LapMCC algorithm. Particularly, we conduct four experiments which are handwritten digits classification on USPS database [30], face recognition on a combinatorial subset selected from Yale-B and Extend Yale-B databases [31], a more challenging face dataset, i.e., ChokePoint dataset [32], and object classification on ETH-80 database [33].

The

Conclusion

Canonical correlation analysis has proven a successful technique for two-view dimension reduction in both theoretical foundation and practical applications. However, the traditional CCA cannot reveal the nonlinear correlation relationship between multiview features. In this paper, we propose Hessian multiset canonical correlations to tackle this disadvantage of CCA. Hessian has richer null space and can properly exploit the intrinsic local geometry of the data manifold in contrast to Laplacian.

Acknowledgement

This study was supported by the National Natural Science Foundation of China under Grants 61671480, 61572486; the Fundamental Research Funds for the Central Universities under Grant 14CX02203A and the Graduate Student Innovation Project Funding of China University of Petroleum (East China) under Grants YCXJ2016075.

References (33)

  • F. Korn et al.

    On the ``dimensionality curse'' and the ``self-similarity blessing''

    IEEE Trans. Knowl. Data Eng.

    (2001)
  • D.P. Foster et al.

    Multi-View Dimensionality Reduction Via Canonical Correlation Analysis

    (2008)
  • Y. Han et al.

    Sparse unsupervised dimensionality reduction for multiple view data

    IEEE Trans. Circuits Syst. Video Technol.

    (2012)
  • H. Hotelling

    Relations between two sets of variates

    Biometrika

    (1936)
  • T.-K. Kim et al.

    Discriminative learning and recognition of image set classes using canonical correlations

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2007)
  • S. Sun et al.

    Multiview uncorrelated discriminant analysis

    IEEE Trans. Cybernet.

    (2016)
  • Cited by (92)

    View all citing articles on Scopus
    View full text