DOCSLIB.ORG

Principal Component Regression with Semirandom Observations Via Matrix Completion

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Principal Component Regression with Semirandom Observations Via Matrix Completion Principal Component Regression with Semirandom Observations via Matrix Completion Aditya Bhaskara Kanchana Ruwanpathirana Maheshakya Wijewardena University of Utah University of Utah University of Utah Abstract variables. Denoting the outputs by a vector y 2 Rn and the prediction variables by a matrix M 2 Rn×d (each row corresponds to an observation), linear re- Principal Component Regression (PCR) is a gression aims to model y as popular method for prediction from data, and is one way to address the so-called multi- y = Mβ + η; (1) collinearity problem in regression. It was shown recently that algorithms for PCR such where β 2 d the vector of regression coefficients, and as hard singular value thresholding (HSVT) R η is a vector of error terms, typically modeled as having are also quite robust, in that they can han- independent Gaussian entries. The goal of the regres- dle data that has missing or noisy covariates. sion problem is to find the coefficients β. The stan- However, such spectral approaches require dard procedure is to solve the least squares problem, strong distributional assumptions on which minky − Mβk2. entries are observed. Specifically, every co- variate is assumed to be observed with prob- One of the issues in high dimensional regression (large ability (exactly) p, for some value of p. Our d) is the problem of multi-collinearity, where there goal in this work is to weaken this require- are correlations (in values) between regression vari- ment, and as a step towards this, we study a ables, leading to unstable values for the parameters \semi-random" model. In this model, every β (see Gunst and Webster (1975); Jolliffe (1986)). covariate is revealed with probability p, and One common approach to overcome this problem is then an adversary comes in and reveals ad- to use subset selection methods (e.g., Hocking (1972); ditional covariates. While the model seems Draper and Smith (1966)). Another classic approach intuitively easier, it is well known that algo- Hotelling (1957); Jolliffe (1986) is to perform regres- rithms such as HSVT perform poorly. Our sion after projecting to the space of the top prin- approach is based on studying the closely re- cipal components of the matrix M. This is called lated problem of Noisy Matrix Completion in Principal Component Regression (PCR). Formally, if a semi-random setting. By considering a new M (r) = U (r)Σ(r)(V (r))T is the best rank r approxima- semidefinite programming relaxation, we de- tion of M, then the idea is to replace (1) with velop new guarantees for matrix completion, which is our core technical contribution. y = MV (r)β0 + η: (2) The goal is to now find the best β0 2 Rr. This is known 1 Introduction to yield more stable results in many settings (Mosteller and Tukey, 1977). It also has the advantage of being Regression is one of the fundamental problems in a \smaller" problem, thus it can be solved more ef- statistics and data analysis, with over two hundred ficiently. Recently, Agarwal et al. (2019) observed years of history. We are given n observations, each that in settings where the regression matrix is (close consisting of d prediction or regression variables and to) low rank, PCR also provides a way to deal with one output that is a linear function of the prediction missing and noisy covariates | a well-known issue in applications (Little, 1992). Their main insight is that Proceedings of the 24th International Conference on Artifi- observing a random subset of the entries of M is good cial Intelligence and Statistics (AISTATS) 2021, San Diego, enough to obtain the top singular vectors V (r) to a California, USA. PMLR: Volume 130. Copyright 2021 by high accuracy. Thus one can obtain a β0 that yields a the author(s). low-error regression model. Principal Component Regression with Semirandom Observations via Matrix Completion While the result of Agarwal et al. (2019) provides novel matrix and applying PCR. Further, matrix completion theoretical guarantees on regression with missing en- can also be formulated (and indeed has been studied tries, it is restrictive: guarantees can only be obtained in Cheng and Ge (2018)) in a semi-random setting. when every entry of M is observed independently with However, to the best of our knowledge, trade-offs be- (the same) probability p. Given inherent dependencies tween the error in the observations and the recovery in applications (e.g., some users may intrinsically re- error have not been studied in a semi-random model. veal more information than others; certain covariates Analyzing this via a semidefinite programming (SDP) may be easier to measure than others, etc.), we ask: relaxation is our main technical contribution. Interest- Can we design algorithms for partially-observed PCR ingly, our analysis relies heavily on a recent non-convex under milder assumptions on the observations? approach to matrix completion (Chen et al., 2019). It is easy to see that unless sufficiently many entries are observed from each column, recovering the corre- 1.1 Our results sponding coordinate of β is impossible. This motivates having a lower bound on the probability of each entry We discuss now our results about matrix completion being revealed. This leads us to considering a so-called and the implications to PCR with missing and noisy semi-random model (Blum and Spencer, 1995; Feige entries. Formal definitions and the full setup is de- and Krauthgamer, 2000; Makarychev et al., 2012; ferred to Section 2. Cheng and Ge, 2018). In our context, such a model Result about matrix completion. Recall that in corresponds to the following: (a) initially, all the en- matrix completion, we have an unknown rank-r ma- tries are revealed independently with probability p, (b) trix M ∗ (dimensions n × n), which we wish to infer an adversary reveals additional elements (arbitrarily). given noisy and partial observations. Ω denotes the Despite appearing to make the problem \easier" than e set of observed indices, and σ denotes the standard the random case, since an algorithm has no idea if an deviation of the noise added to each observation. Ω is entry was revealed in step (a) or (b), algorithms based e chosen using a semi-random process (first forming Ω on obtaining estimators (that rely on entries being ob- in which every index is present with probability p and served with probability p) will fail. then adversarially adding indices). Theorem (informal). Under appropriate incoher- Further motivation for the semirandom model. ∗ In both PCR with missing covariates and matrix com- ence assumptions on M , there exists a polynomial pletion, at a high level, we wish to understand what time algorithm that finds an estimate Z that satisfies \observation patterns" allow for effective recovery. kM − ZkF ≤ Oκ,p,µ (nr log n · σ) : Without stochasticity assumptions, there are hardness 2 results (even when one observes Ω(n ) entries; Hardt The result holds with high probability, and the formal et al. (2014)), and we know of positive results when statement along with conditions for p; σ is presented in each entry is observed independently with probability Theorem 1. The parameters κ, µ capture the condition p. The semirandom model has been studied in the number and incoherence properties of M ∗. To the best literature (for graph partitioning and also for matrix of our knowledge, such a bound relating the noise and completion) because in spite of seeming easier as dis- recovery error is not previously known in the semi- cussed above, it ends up causing spectral methods to random model for matrix completion. The prior work fail. The semirandom model is equivalent to the set- of (Cheng and Ge, 2018) does not consider the noisy ting where every entry has some \base" probability of case. being observed, but the probability could be higher for p some (unknown to us) entries. E.g., in recommender Our recovery error bound is a factor n worse than systems, some users may provide more ratings for cer- the best (and optimal) bounds for matrix completion tain types of items than others, and this is typically with random observations (Chen et al., 2019; Kesha- unknown to the algorithm. The model is also closely van et al., 2012). However, it is better than some of related to the notion of Massart noise which is known the earlier results of (Candes and Plan, 2009). Specif- to be challenging for a variety of learning problems ically, the work of (Candes and Plan, 2009) imposes (see, e.g., Diakonikolas et al. (2019)). the constraint kPΩ(Z − M)kF ≤ δ and shows that q kZ − M ∗k ≤ n δ (up to constants). If Ω corre- Our approach to partially-observed PCR will be based F p sponds to i.i.d. observations with probability p, we on the closely related problem of matrix completion p (see Candes and Recht (2008); Keshavan et al. (2012); must set δ = n p · σ for the SDP to be feasible. This Bhojanapalli and Jain (2014) and references therein). leads to a recovery error bound of n3=2σ. In the semi- The problems are related because intuitively, we can random model, the SDP feasibility constraint must think of filling in the missing values in the covariate now depend on Ωe and if we have n2p observations, Aditya Bhaskara, Kanchana Ruwanpathirana, Maheshakya Wijewardena the n3=2σ bound becomes even worse. Chen et al. 1.2 Related work (2019) improve the error bound above by a factor of p n p in the random case. Our result can be viewed There has been extensive literature on PCR and ma- as being in between the two results while holding true trix completion that we do not review for lack of space.
Load more

Recommended publications
  • Survey on Probabilistic Models of Low-Rank Matrix Factorizations
    Article Survey on Probabilistic Models of Low-Rank Matrix Factorizations Jiarong Shi 1,2,*, Xiuyun Zheng 1 and Wei Yang 1 1 School of Science, Xi’an University of Architecture and Technology, Xi’an 710055, China; [email protected] (X.Z.); [email protected] (W.Y.) 2 School of Architecture, Xi’an University of Architecture and Technology, Xi’an 710055, China * Correspondence: [email protected]; Tel.: +86-29-8220-5670 Received: 12 June 2017; Accepted: 16 August 2017; Published: 19 August 2017 Abstract: Low-rank matrix factorizations such as Principal Component Analysis (PCA), Singular Value Decomposition (SVD) and Non-negative Matrix Factorization (NMF) are a large class of methods for pursuing the low-rank approximation of a given data matrix. The conventional factorization models are based on the assumption that the data matrices are contaminated stochastically by some type of noise. Thus the point estimations of low-rank components can be obtained by Maximum Likelihood (ML) estimation or Maximum a posteriori (MAP). In the past decade, a variety of probabilistic models of low-rank matrix factorizations have emerged. The most significant difference between low-rank matrix factorizations and their corresponding probabilistic models is that the latter treat the low-rank components as random variables. This paper makes a survey of the probabilistic models of low-rank matrix factorizations. Firstly, we review some probability distributions commonly-used in probabilistic models of low-rank matrix factorizations and introduce the conjugate priors of some probability distributions to simplify the Bayesian inference. Then we provide two main inference methods for probabilistic low-rank matrix factorizations, i.e., Gibbs sampling and variational Bayesian inference.
    [Show full text]
  • Solving a Low-Rank Factorization Model for Matrix Completion by a Nonlinear Successive Over-Relaxation Algorithm
    SOLVING A LOW-RANK FACTORIZATION MODEL FOR MATRIX COMPLETION BY A NONLINEAR SUCCESSIVE OVER-RELAXATION ALGORITHM ZAIWEN WEN †, WOTAO YIN ‡, AND YIN ZHANG § Abstract. The matrix completion problem is to recover a low-rank matrix from a subset of its entries. The main solution strategy for this problem has been based on nuclear-norm minimization which requires computing singular value decompositions – a task that is increasingly costly as matrix sizes and ranks increase. To improve the capacity of solving large-scale problems, we propose a low-rank factorization model and construct a nonlinear successive over-relaxation (SOR) algorithm that only requires solving a linear least squares problem per iteration. Convergence of this nonlinear SOR algorithm is analyzed. Numerical results show that the algorithm can reliably solve a wide range of problems at a speed at least several times faster than many nuclear-norm minimization algorithms. Key words. Matrix Completion, alternating minimization, nonlinear GS method, nonlinear SOR method AMS subject classifications. 65K05, 90C06, 93C41, 68Q32 1. Introduction. The problem of minimizing the rank of a matrix arises in many applications, for example, control and systems theory, model reduction and minimum order control synthesis [20], recovering shape and motion from image streams [25, 32], data mining and pattern recognitions [6] and machine learning such as latent semantic indexing, collaborative prediction and low-dimensional embedding. In this paper, we consider the Matrix Completion (MC) problem of finding a lowest-rank matrix given a subset of its entries, that is, (1.1) min rank(W ), s.t. Wij = Mij , ∀(i,j) ∈ Ω, W ∈Rm×n where rank(W ) denotes the rank of W , and Mi,j ∈ R are given for (i,j) ∈ Ω ⊂ {(i,j) : 1 ≤ i ≤ m, 1 ≤ j ≤ n}.
    [Show full text]
  • Riemannian Gradient Descent Methods for Graph-Regularized Matrix Completion
    Riemannian Gradient Descent Methods for Graph-Regularized Matrix Completion Shuyu Dongy, P.-A. Absily, K. A. Gallivanz∗ yICTEAM Institute, UCLouvain, Belgium zFlorida State University, U.S.A. December 2, 2019 Abstract Low-rank matrix completion is the problem of recovering the missing entries of a data matrix by using the assumption that the true matrix admits a good low-rank approxi- mation. Much attention has been given recently to exploiting correlations between the column/row entities to improve the matrix completion quality. In this paper, we propose preconditioned gradient descent algorithms for solving the low-rank matrix completion problem with graph Laplacian-based regularizers. Experiments on synthetic data show that our approach achieves significant speedup compared to an existing method based on alternating minimization. Experimental results on real world data also show that our methods provide low-rank solutions of similar quality in comparable or less time than the state-of-the-art method. Keywords: matrix completion; graph information; Riemannian optimization; low-rank op- timization MSC: 90C90, 53B21, 15A83 1 Introduction Low-rank matrix completion arises in applications such as recommender systems, forecasting and imputation of data; see [38] for a recent survey. Given a data matrix with missing entries, the objective of matrix completion can be formulated as the minimization of an error function of a matrix variable with respect to the data matrix restricted to its revealed entries. In various applications, the data matrix either has a rank much lower than its dimensions or can be approximated by a low-rank matrix. As a consequence, restricting the rank of the matrix variable in the matrix completion objective not only corresponds to a reasonable assumption for successful recovery of the data matrix but also reduces the model complexity.
    [Show full text]
  • Low-Rank Matrix Completion Yuejie Chi, Senior Member, IEEE
    1 Low-Rank Matrix Completion Yuejie Chi, Senior Member, IEEE Imagine one observes a small subset of entries in a large matrix and aims to recover the entire matrix. Without a priori knowledge of the matrix, this problem is highly ill-posed. Fortunately, data matrices often exhibit low- dimensional structures that can be used effectively to regularize the solution space. The celebrated effectiveness of Principal Component Analysis (PCA) in science and engineering suggests that most variability of real-world data can be accounted for by projecting the data onto a few directions known as the principal components. Correspondingly, the data matrix can be modeled as a low-rank matrix, at least approximately. Is it possible to complete a partially observed matrix if its rank, i.e., its maximum number of linearly-independent row or column vectors, is small? Low-rank matrix completion arises in a variety of applications in recommendation systems, computer vision, and signal processing. As a motivating example, consider users’ ratings of products arranged in a rating matrix. Each rating may only be affected by a small number of factors – such as price, quality, and utility – and how they are reflected on the products’ specifications and users’ expectations. Naturally, this suggests that the rating matrix is low-rank, since the numbers of users and products are much higher than the number of factors. Often, the rating matrix is sparsely observed, and it is of great interest to predict the missing ratings to make targeted recommendations. I. RELEVANCE The theory and algorithms of low-rank matrix completion have been significantly expanded in the last decade with converging efforts from signal processing, applied mathematics, statistics, optimization and machine learning.
    [Show full text]
  • Low-Rank Matrix Completion Via Preconditioned Optimization on the Grassmann Manifold
    Low-rank matrix completion via preconditioned optimization on the Grassmann manifold Nicolas Boumala, P.-A. Absilb aInria & D.I., UMR 8548, Ecole Normale Sup´erieure, Paris, France. bDepartment of Mathematical Engineering, ICTEAM Institute, Universit´ecatholique de Louvain, Belgium. Abstract We address the numerical problem of recovering large matrices of low rank when most of the entries are unknown. We exploit the geometry of the low-rank constraint to recast the problem as an unconstrained optimization problem on a single Grassmann manifold. We then apply second-order Riemannian trust-region methods (RTRMC 2) and Riemannian conjugate gradient methods (RCGMC) to solve it. A preconditioner for the Hessian is introduced that helps control the conditioning of the problem and we detail preconditioned versions of Riemannian optimization algorithms. The cost of each iteration is linear in the number of known entries. The proposed methods are competitive with state-of-the-art algorithms on a wide range of problem instances. In particular, they perform well on rectangular matrices. We further note that second-order and preconditioned methods are well suited to solve badly conditioned matrix completion tasks. Keywords: Low-rank matrix completion, optimization on manifolds, Grassmann manifold, preconditioned Riemannian trust-regions, preconditioned Riemannian conjugate gradient, second-order methods, fixed-rank geometry, RTRMC, RCGMC 2000 MSC: 90C90, 53B21, 15A83 1. Introduction This paper considers the problem of recovering a low-rank m × n matrix X (m ≤ n) of which a few entries are observed, possibly with noise. Throughout, we assume that r = rank(X) is known and is much smaller than m, and we let Ω ⊂ f1 : : : mg × f1 : : : ng be the set of indices of the observed entries of X, that is, Xij is known if (i; j) is in Ω.
    [Show full text]
  • Faster Matrix Completion Using Randomized SVD
    Faster Matrix Completion Using Randomized SVD Xu Feng Wenjian Yu Yaohang Li BNRist, Dept. Computer Science & Tech. BNRist, Dept. Computer Science & Tech. Dept. Computer Science Tsinghua University Tsinghua University Old Dominion University Beijing, China Beijing, China Norfolk, VA 23529, USA [email protected] [email protected] [email protected] Abstract—Matrix completion is a widely used technique for handling large data set, because the singular values exceeding image inpainting and personalized recommender system, etc. a threshold and the corresponding singular vectors need to In this work, we focus on accelerating the matrix completion be computed in each iteration step. Truncated singular value using faster randomized singular value decomposition (rSVD). Firstly, two fast randomized algorithms (rSVD-PI and rSVD- decomposition (SVD), implemented with svds in Matlab or BKI) are proposed for handling sparse matrix. They make use of lansvd in PROPACK [5], is usually employed in the SVT an eigSVD procedure and several accelerating skills. Then, with algorithm [1]. Another method for matrix completion is the the rSVD-BKI algorithm and a new subspace recycling technique, inexact augmented Lagrange multiplier (IALM) algorithm [6], we accelerate the singular value thresholding (SVT) method in [1] which also involves singular value thresholding and was orig- to realize faster matrix completion. Experiments show that the proposed rSVD algorithms can be 6X faster than the basic rSVD inally proposed for the robust principal component analysis algorithm [2] while keeping same accuracy. For image inpainting (PCA) problem [7]. With artificially-generated low-rank ma- and movie-rating estimation problems (including up to 2 × 107 trices, experiments in [6] demonstrated that IALM algorithm ratings), the proposed accelerated SVT algorithm consumes 15X could be several times faster than the SVT algorithm.
    [Show full text]
  • High-Dimensional Statistical Inference: from Vector to Matrix
    University of Pennsylvania ScholarlyCommons Publicly Accessible Penn Dissertations 2015 High-dimensional Statistical Inference: from Vector to Matrix Anru Zhang University of Pennsylvania, [email protected] Follow this and additional works at: https://repository.upenn.edu/edissertations Part of the Applied Mathematics Commons, and the Statistics and Probability Commons Recommended Citation Zhang, Anru, "High-dimensional Statistical Inference: from Vector to Matrix" (2015). Publicly Accessible Penn Dissertations. 1172. https://repository.upenn.edu/edissertations/1172 This paper is posted at ScholarlyCommons. https://repository.upenn.edu/edissertations/1172 For more information, please contact [email protected]. High-dimensional Statistical Inference: from Vector to Matrix Abstract Statistical inference for sparse signals or low-rank matrices in high-dimensional settings is of significant interest in a range of contemporary applications. It has attracted significant ecentr attention in many fields including statistics, applied mathematics and electrical engineering. In this thesis, we consider several problems in including sparse signal recovery (compressed sensing under restricted isometry) and low-rank matrix recovery (matrix recovery via rank-one projections and structured matrix completion). The first part of the thesis discusses compressed sensing and affineank r minimization in both noiseless and noisy cases and establishes sharp restricted isometry conditions for sparse signal and low-rank matrix recovery. The analysis
    [Show full text]
  • Matrix Completion with Queries
    Matrix Completion with Queries Natali Ruchansky Mark Crovella Evimaria Terzi Boston University Boston University Boston University [email protected] [email protected] [email protected] ABSTRACT the matrix contains the volumes of traffic among source- In many applications, e.g., recommender systems and traffic destination pairs), and computer vision (image matrices). monitoring, the data comes in the form of a matrix that is In many cases, only a small percentage of the matrix entries only partially observed and low rank. A fundamental data- are observed. For example, the data used in the Netflix prize analysis task for these datasets is matrix completion, where competition was a matrix of 480K users × 18K movies, but the goal is to accurately infer the entries missing from the only 1% of the entries were known. matrix. Even when the data satisfies the low-rank assump- A common approach for recovering such missing data is tion, classical matrix-completion methods may output com- called matrix completion. The goal of matrix-completion pletions with significant error { in that the reconstructed methods is to accurately infer the values of missing entries, matrix differs significantly from the true underlying ma- subject to certain assumptions about the complete matrix [2, trix. Often, this is due to the fact that the information 3, 5, 7, 8, 9, 19]. For a true matrix T with observed val- contained in the observed entries is insufficient. In this work, ues only in a set of positions Ω, matrix-completion methods we address this problem by proposing an active version of exploit the information in the observed entries in T (de- matrix completion, where queries can be made to the true noted TΩ) in order to produce a \good" estimate Tb of T .
    [Show full text]
  • Robust Low-Rank Matrix Completion Via an Alternating Manifold Proximal Gradient Continuation Method Minhui Huang, Shiqian Ma, and Lifeng Lai
    1 Robust Low-rank Matrix Completion via an Alternating Manifold Proximal Gradient Continuation Method Minhui Huang, Shiqian Ma, and Lifeng Lai Abstract—Robust low-rank matrix completion (RMC), or ro- nonconvex. To resolve this issue, there is a line of research bust principal component analysis with partially observed data, focusing on convex relaxation of RMC formulations [5], [6]. has been studied extensively for computer vision, signal process- For example, a widely used technique in these works is ing and machine learning applications. This problem aims to decompose a partially observed matrix into the superposition of employing its tightest convex proxy, i:e: the nuclear norm. a low-rank matrix and a sparse matrix, where the sparse matrix However, calculating the nuclear norm in each iteration can captures the grossly corrupted entries of the matrix. A widely bring numerical difficulty for large-scale problems in practice, used approach to tackle RMC is to consider a convex formulation, because algorithms for solving them need to compute a singular which minimizes the nuclear norm of the low-rank matrix (to value decomposition (SVD) in each iteration, which can be promote low-rankness) and the `1 norm of the sparse matrix (to promote sparsity). In this paper, motivated by some recent works very time consuming. Recently, nonconvex RMC formulations on low-rank matrix completion and Riemannian optimization, we based on low-rank matrix factorization were proposed in the formulate this problem as a nonsmooth Riemannian optimization literature [7], [8], [9], [10]. In these formulations, the target problem over Grassmann manifold. This new formulation is matrix is factorized as the product of two much smaller matrices scalable because the low-rank matrix is factorized to the mul- so that the low-rank property is automatically satisfied.
    [Show full text]
  • Rank-One Matrix Pursuit for Matrix Completion
    Rank-One Matrix Pursuit for Matrix Completion Zheng Wang∗ [email protected] Ming-Jun Laiy [email protected] Zhaosong Luz [email protected] Wei Fanx [email protected] Hasan Davulcu{ [email protected] Jieping Ye∗¶ [email protected] ∗The Biodesign Institue, Arizona State University, Tempe, AZ 85287, USA yDepartment of Mathematics, University of Georgia, Athens, GA 30602, USA zDepartment of Mathematics, Simon Fraser University, Burnaby, BC, V5A 156, Canada xHuawei Noah’s Ark Lab, Hong Kong Science Park, Shatin, Hong Kong {School of Computing, Informatics, and Decision Systems Engineering, Arizona State University, Tempe, AZ 85281, USA Abstract 1. Introduction Low rank matrix completion has been applied Low rank matrix learning has attracted significant attention successfully in a wide range of machine learn- in the machine learning community due to its wide range ing applications, such as collaborative filtering, of applications, such as collaborative filtering (Koren et al., image inpainting and Microarray data imputa- 2009; Srebro et al., 2005), compressed sensing (Candes` tion. However, many existing algorithms are not & Recht, 2009), multi-class learning and multi-task learn- scalable to large-scale problems, as they involve ing (Argyriou et al., 2008; Negahban & Wainwright, 2010; computing singular value decomposition. In this Dud´ık et al., 2012). In this paper, we consider the general paper, we present an efficient and scalable algo- form of low rank matrix completion: given a partially ob- rithm for matrix completion. The key idea is to served real-valued matrix Y 2 <n×m, the low rank matrix extend the well-known orthogonal matching pur- completion problem is to find a matrix X 2 <n×m with suit from the vector case to the matrix case.
    [Show full text]
  • Leave-One-Out Approach for Matrix Completion: Primal and Dual Analysis,” Arxiv Preprint Arxiv:1803.07554V1, 2018
    Leave-one-out Approach for Matrix Completion: Primal and Dual Analysis Lijun Ding and Yudong Chen∗ Abstract In this paper, we introduce a powerful technique based on Leave-One-Out analysis to the study of low-rank matrix completion problems. Using this technique, we develop a general approach for obtaining fine-grained, entrywise bounds for iterative stochastic procedures in the presence of probabilistic depen- dency. We demonstrate the power of this approach in analyzing two of the most important algorithms for matrix completion: (i) the non-convex approach based on Projected Gradient Descent (PGD) for a rank-constrained formulation, also known as the Singular Value Projection algorithm, and (ii) the convex relaxation approach based on nuclear norm minimization (NNM). Using this approach, we establish the first convergence guarantee for the original form of PGD without regularization or sample splitting, and in particular shows that it converges linearly in the infinity norm. For NNM, we use this approach to study a fictitious iterative procedure that arises in the dual analysis. Our results show that NNM recovers an d-by-d rank-r matrix with (µr log(µr)d log d) observed entries. This bound has optimal dependence on the matrix dimension and is indeO pendent of the condition number. To the best of our knowledge, none of previous sample complexity results for tractable matrix completion algorithms satisfies these two properties simultaneously. 1 Introduction The matrix completion problem concerns recovering a low-rank matrix given a (typically random) subset of its entries. To study the sample complexity and algorithmic behaviors of this problem, one often needs to analyze an iterative procedure in the presence of dependency across the iterations and the entries of the iterates.
    [Show full text]
  • A New Theory for Matrix Completion
    A New Theory for Matrix Completion Guangcan Liu∗ Qingshan Liuy Xiao-Tong Yuanz B-DAT, School of Information & Control, Nanjing Univ Informat Sci & Technol NO 219 Ningliu Road, Nanjing, Jiangsu, China, 210044 {gcliu,qsliu,xtyuan}@nuist.edu.cn Abstract Prevalent matrix completion theories reply on an assumption that the locations of the missing data are distributed uniformly and randomly (i.e., uniform sampling). Nevertheless, the reason for observations being missing often depends on the unseen observations themselves, and thus the missing data in practice usually occurs in a nonuniform and deterministic fashion rather than randomly. To break through the limits of random sampling, this paper introduces a new hypothesis called isomeric condition, which is provably weaker than the assumption of uniform sampling and arguably holds even when the missing data is placed irregularly. Equipped with this new tool, we prove a series of theorems for missing data recovery and matrix completion. In particular, we prove that the exact solutions that identify the target matrix are included as critical points by the commonly used nonconvex programs. Unlike the existing theories for nonconvex matrix completion, which are built upon the same condition as convex programs, our theory shows that nonconvex programs have the potential to work with a much weaker condition. Comparing to the existing studies on nonuniform sampling, our setup is more general. 1 Introduction Missing data is a common occurrence in modern applications such as computer vision and image processing, reducing significantly the representativeness of data samples and therefore distorting seriously the inferences about data. Given this pressing situation, it is crucial to study the problem of recovering the unseen data from a sampling of observations.
    [Show full text]