Weighted Schatten -norm minimization with logarithmic constraint for subspace clustering☆
Introduction
Subspace clustering has significantly facilitated image processing tasks such as face clustering [1] and motion segmentation [2], [32], whose purpose is to explore the multiple underlying low-dimensional subspaces and segment the high-dimensional data into these relative low-dimensional subspaces. Over the past few decades, numerous subspace clustering algorithms have been extensively studied. Most of them can roughly fall into four categories: factorization-based methods [5]; algebraic methods [3], [4]; statistical methods [6]; self-expressive methods [8], [9], [10], [11]. In this paper, we focus on the self-expressive methods owing to its promising property.
In self-expressive methods, each data sample drawn from a union of low-dimensional subspaces can be linearly represented by samples in the same subspace. To learn a more discriminative affinity matrix, the self-expressive methods commonly employ the sparse or low-rank prior as the regularization of the coefficient matrix. For example, the SSC [7] get a sparse coefficient matrix, which utilize the norm. However, the coefficient matrix in SSC may be too sparse to poor result, because it uses only the local structure. To overcome this, many self-expressive methods based on low-rank representation have been widely developed, which integrate the global information, such as LRR [10], FLRR [22], WNNM [34]. Instead of directly solving the rank function with NP-hard complexity, these low-rank methods replace it with the nuclear norm, which is the convex envelope of the rank function over the unit ball of matrices [12]. However, the results of the nuclear-norm minimization may deviate from that of the real rank minimization. That is because the real rank minimization considers that the contributions of all nonzero singular values of the coefficient matrix are equal, while the nuclear-norm minimization simply adds all the nonzero singular values [13].
To conquer the above issue, some non-convex low-rank surrogates [14], [15], [16], [17] have been developed to further replace the nuclear norm. These low-rank methods treat the nonzero singular values in a soft way, where the contributions of the large singular values are declined compared with the nuclear-norm minimization. For example, the Schatten- norm [15], [16], [18] and the logarithm of the determinant [13], [19], [20] are two representative non-convex low-rank surrogates. In detail, Schatten- norm achieves to apply more suppression on the large singular values. By adjusting the parameter , a family of non-convex low-rank surrogates can be obtained, which are between the nuclear norm and real rank function. For example, the Schatten- norm can be treated as the rank function at . Nevertheless, the Schatten- norm will promote the contribution of the small singular values that are generally regarded as noise. As for the logarithm of the determinant, it not only can reduce the contribution of large singular values, but also urge the small singular values to zero. However, the logarithm of the determinant, as one non-convex low-rank surrogate, can not infinitely approximate the rank function. Besides, these non-convex low-rank surrogates may yield unsatisfactory results, because they ignore that the singular values of the coefficient matrix can be suppressed to a greater degree.
Motivated by the advantages and weaknesses of the Schatten- norm and the logarithm of the determinant, we propose two novel non-convex low-rank surrogates named SLog and WSLog. That is, our methods can combine the advantages of the two previous methods and make up for their weaknesses at the same time. On the other hand, combining the Schatten-p norm and the logarithm of the determinant can suppress the small and large singular values flexibly with a greater degree. Specifically, we firstly proposed SLog by utilizing the logarithmic to tighten the Schatten- norm. As shown in Fig. 1(a), our method SLog can approximate the rank function more closely in comparison with other surrogates. Meanwhile, to suppress the small singular values simultaneously, we propose a weighted Schatten -norm minimization named WSLog based on SLog by introducing the weight . Our method WSLog can shrink the small singular values to a greater degree, especially when . Extensive experimental results demonstrate that the two proposed methods achieve comparable and even better accuracy. Our contributions are summarized as follows:
1) We firstly propose SLog by utilizing the logarithmic to tighten the Schatten- norm, which can shrink the large singular values in a similar trend of the real rank minimization, and approximate the rank function more closely.
2) To shrink the small singular values to a greater degree simultaneously, we propose a weighted Schatten -norm minimization named WSLog based on SLog by introducing the weight to mitigate the negative contribution of noise, especially .
3) We present an optimization algorithm to solve SLog and WSLog problems. Extensive experiments demonstrate that the proposed methods outperform other methods across most of datasets.
Section snippets
Related works
Given a data matrix , denotes the number of samples and each sample is represented by an -dimensional column vector (i.e., ). The low-rank problem for subspace clustering can be written aswhere the low-rank representation denotes the coefficient matrix. is introduced as sparse noise, which is often regularized by norm, such as , where is the th column of . represents the balanced parameter
Construction of SLog
In fact, the existing methods of non-convex low-rank approximation including Schatten- norm and the logarithm of the determinant ignore that the large singular values of the coefficient matrix should be suppressed with a greater degree, which may not be capable of exploiting the subspace structures of data faithfully, and hence may produce the unsatisfactory results. Thus, there is a need to design a closer surrogate function of the real rank function, which is expected to suppress these large
Connection to related works
Recently, Peng et al. [13] proposed the Log-det, which is written asIn problem (20), the data matrix , where is the unknown underlying clean data matrix, denotes the sparse error, and represents the Gaussian noise. denotes the low-rank representation, and is the th singular value of . denotes a proper norm to handle sparse noise or outliers. and are two positive balancing parameters.
Connection to Log-det: Both
Experiments
In this section, we conduct extensive experiments to evaluate the effectiveness of the proposed methods, which are conducted on an Intel Core i5-8500 workstation with 3.00-GHz CPU and 16-GB memory. In these experiments, we compare our methods with the classical LRR methods based on nuclear norm (i.e., LRR [10] and FLRR [22], where FLRR is a fast version of LRR.), and the non-convex low-rank approximations (i.e., Schatten-norm [15], [16], FULRR [11], RMC-NC [20], and Log-det [13]). For the
Conclusions
In this paper, we proposed a novel non-convex low-rank approximation by applying the weighted Schatten- norm with logarithmic constraint, which can suppress the small and large singular values flexibly with a greater degree, and be more powerful to exploit the subspace structures. Specifically, we firstly proposed SLog by utilizing the logarithmic to tighten the Schatten- norm, which can shrink the large singular values in a similar trend of the real rank minimization. Furthermore, to shrink
CRediT authorship contribution statement
Qiangqiang Shen: Writing – original draft, Conceptualization, Methodology, Software. Yongyong Chen: Writing – review & editing, Validation, Visualization. Yongsheng Liang: Supervision. Shuangyan Yi: Data curation, Software. Wei Liu: Software, Investigation.
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.
References (34)
- et al.
Scalable and robust sparse subspace clustering using randomized clustering and multilayer graphs
Signal Process.
(2019) - et al.
Multiple kernel low-rank representation-based robust multi-view subspace clustering
Inf. Sci.
(2021) - et al.
Efficient sparse subspace clustering by nearest neighbour filtering
Signal Process.
(2020) - et al.
Provable subspace clustering: when LRR meets SSC
IEEE Trans. Image Process.
(2019) - et al.
Shape interaction matrix revisited and robustified: efficient subspace clustering with corrupted and incomplete data
Proc. ICCV
(2015) - et al.
Estimation of subspace arrangements with applications in modeling and segmenting mixed data
SIAM Rev.
(2008) - et al.
Multibody factorization with uncertainty and missing data using the em algorithm
Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
(2004) - et al.
Motion segmentation in the presence of outlying, incomplete, or corrupted trajectories
IEEE Trans. Pattern Anal. Mach. Intell.
(2010) - et al.
Sparse subspace clustering: algorithm, theory, and applications
IEEE Trans. Pattern Anal. Mach. Intell.
(2013) - et al.
Efficient solvers for sparse subspace clustering
Signal Process.
(2020)
Robust subspace segmentation by low-rank representation
Proc. 27rd Int. Conf. Mach. Learn
Fast universal low rank representation
IEEE Trans. Circuits Syst. Video Technol.
Matrix rank minimization with applications
Subspace clustering using log-determinant rank approximation
Proc. 19th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining.
Large-scale low-rank matrix learning with nonconvex regularizers
IEEE Trans. Pattern Anal. Mach. Intell.
Minimal shrinkage for noisy data recovery using Schatten- norm objective
Proc. ECML PKDD
Low-rank matrix recovery via efficient Schatten p-norm minimization
Proc. AAAI Conf. Artif. Intell.
Cited by (8)
Quaternion tensor completion with sparseness for color video recovery
2024, Applied Soft ComputingLatLRR for subspace clustering via reweighted Frobenius norm minimization
2023, Expert Systems with ApplicationsUnified Framework for Faster Clustering via Joint Schatten p-Norm Factorization With Optimal Mean
2024, IEEE Transactions on Neural Networks and Learning SystemsImage Segmentation Based on Fuzzy Low-Rank Structural Clustering
2023, IEEE Transactions on Fuzzy Systems
- ☆
This work was supported in part by the National Natural Science Foundation of China under Grant No. 62031013, No. 61871154, and No. 62106063, Shenzhen College Stability Support Plan under Grant No. GXWD20201230155427003-20200824113231001, Youth Program of National Natural Science Foundation of China under Grant No. 61906124, and Basic and Applied Basic Research Fund of Guangdong Province under Grant No. 2019A1515011307.