Elizaveta Levina

Regularized estimation of large covariance matrices

This paper considers estimating a covariance matrix of p variables from n observations by either banding the sample covariance matrix or estimating a banded version of the inverse of the covariance. We show that these estimates are consistent in the operator norm as long as (logp) 2 =n ! 0, and obtain explicit rates. The results are uniform over some fairly natural well-conditioned families of covariance matrices. We also introduce an analogue of the Gaussian white noise model and show that if the population covariance is embeddable in that model and well-conditioned then the banded approximations produce consistent estimates of the eigenvalues and associated eigenvectors of the covariance matrix. The results can be extended to smooth versions of banding and to non-Gaussian distributions with su‐ciently short tails. A resampling approach is proposed for choosing the banding parameter in practice. This approach is illustrated numerically on both simulated and real data.

Covariance regularization by thresholding

This paper considers regularizing a covariance matrix of p variables estimated from n observations, by hard thresholding. We show that the thresholded estimate is consistent in the operator norm as long as the true covariance matrix is sparse in a suitable sense, the variables are Gaussian or sub-Gaussian, and (log p)/n → 0, and obtain explicit rates. The results are uniform over families of covariance matrices which satisfy a fairly natural notion of sparsity. We discuss an intuitive resampling scheme for threshold selection and prove a general cross-validation result that justifies this approach. We also compare thresholding to other covariance estimators in simulations and on an example from climate data. 1. Introduction. Estimation of covariance matrices is important in a number of areas of statistical analysis, including dimension reduction by principal component analysis (PCA), classification by linear or quadratic discriminant analysis (LDA and QDA), establishing independence and conditional independence relations in the context of graphical models, and setting confidence intervals on linear functions of the means of the components. In recent years, many application areas where these tools are used have been dealing with very high-dimensional datasets, and sample sizes can be very small relative to dimension. Examples include genetic data, brain imaging, spectroscopic imaging, climate data and many others. It is well known by now that the empirical covariance matrix for samples of size n from a p-variate Gaussian distribution, Np(μ, � p), is not a good estimator of the population covariance if p is large. Many results in random matrix theory illustrate this, from the classical Mary law [29] to the more recent work of Johnstone and his students on the theory of the largest eigenvalues [12, 23, 30] and associated eigenvectors [24]. However, with the exception of a method for estimating the covariance spectrum [11], these probabilistic results do not offer alternatives to the sample covariance matrix. Alternative estimators for large covariance matrices have therefore attracted a lot of attention recently. Two broad classes of covariance estimators have emerged: those that rely on a natural ordering among variables, and assume that variables

Sparse permutation invariant covariance estimation

The paper proposes a method for constructing a sparse estima- tor for the inverse covariance (concentration) matrix in high-dimensional settings. The estimator uses a penalized normal likelihood approach and forces sparsity by using a lasso-type penalty. We establish a rate of con- vergence in the Frobenius norm as both data dimension p and sample size n are allowed to grow, and show that the rate depends explicitly on how sparse the true concentration matrix is. We also show that a correlation- based version of the method exhibits better rates in the operator norm. We also derive a fast iterative algorithm for computing the estimator, which relies on the popular Cholesky decomposition of the inverse but produces a permutation-invariant estimator. The method is compared to other es- timators on simulated data and on a real data example of tumor tissue classification using gene expression data.

Generalized Thresholding of Large Covariance Matrices

We propose a new class of generalized thresholding operators that combine thresholding with shrinkage, and study generalized thresholding of the sample covariance matrix in high dimensions. Generalized thresholding of the covariance matrix has good theoretical properties and carries almost no computational burden. We obtain an explicit convergence rate in the operator norm that shows the tradeoff between the sparsity of the true model, dimension, and the sample size, and shows that generalized thresholding is consistent over a large class of models as long as the dimension p and the sample size n satisfy log p/n → 0. In addition, we show that generalized thresholding has the “sparsistency” property, meaning it estimates true zeros as zeros with probability tending to 1, and, under an additional mild condition, is sign consistent for nonzero elements. We show that generalized thresholding covers, as special cases, hard and soft thresholding, smoothly clipped absolute deviation, and adaptive lasso, and compare different types of generalized thresholding in a simulation study and in an example of gene clustering from a microarray experiment with tumor tissues.

Robustness of community structure in networks

The discovery of community structure is a common challenge in the analysis of network data. Many methods have been proposed for finding community structure, but few have been proposed for determining whether the structure found is statistically significant or whether, conversely, it could have arisen purely as a result of chance. In this paper we show that the significance of community structure can be effectively quantified by measuring its robustness to small perturbations in network structure. We propose a suitable method for perturbing networks and a measure of the resulting change in community structure and use them to assess the significance of community structure in a variety of networks, both real and computer generated.

Sparse estimation of large covariance matrices via a nested Lasso penalty

The paper proposes a new covariance estimator for large covariance matrices when the variables have a natural ordering. Using the Cholesky decomposition of the inverse, we impose a banded structure on the Cholesky factor, and select the bandwidth adaptively for each row of the Cholesky factor, using a novel penalty we call nested Lasso. This structure has more flexibility than regular banding, but, unlike regular Lasso applied to the entries of the Cholesky factor, results in a sparse estimator for the inverse of the covariance matrix. An iterative algorithm for solving the optimization problem is developed. The estimator is compared to a number of other covariance estimators and is shown to do best, both in simulations and on a real data example. Simulations show that the margin by which the estimator outperforms its competitors tends to increase with dimension.

Sparse Multivariate Regression With Covariance Estimation.

We propose a procedure for constructing a sparse estimator of a multivariate regression coefficient matrix that accounts for correlation of the response variables. This method, which we call multivariate regression with covariance estimation (MRCE), involves penalized likelihood with simultaneous estimation of the regression coefficients and the covariance structure. An efficient optimization algorithm and a fast approximation are developed for computing MRCE. Using simulation studies, we show that the proposed method outperforms relevant competitors when the responses are highly correlated. We also apply the new method to a finance example on predicting asset returns. An R-package containing this dataset and code for computing MRCE and its approximation are available online.

Pseudo-likelihood methods for community detection in large sparse networks

Many algorithms have been proposed for fitting network models with communities, but most of them do not scale well to large networks, and often fail on sparse networks. Here we propose a new fast pseudo-likelihood method for fitting the stochastic block model for networks, as well as a variant that allows for an arbitrary degree distribution by conditioning on degrees. We show that the algorithms perform well under a range of settings, including on very sparse networks, and illustrate on the example of a network of political blogs. We also propose spectral clustering with perturbations, a method of independent interest, which works well on sparse networks where regular spectral clustering fails, and use it to provide an initial value for pseudo-likelihood. We prove that pseudo-likelihood provides consistent estimates of the communities under a mild condition on the starting value, for the case of a block model with two communities.

Consistency of community detection in networks under degree-corrected stochastic block models

Community detection is a fundamental problem in network analysis, with applications in many diverse areas. The stochastic block model is a common tool for model-based community detection, and asymptotic tools for checking consistency of community detection under the block model have been recently developed. However, the block model is limited by its assumption that all nodes within a community are stochastically equivalent, and provides a poor fit to networks with hubs or highly varying node degrees within communities, which are common in practice. The degree-corrected stochastic block model was proposed to address this shortcoming and allows variation in node degrees within a community while preserving the overall block community structure. In this paper we establish general theory for checking consistency of community detection under the degree-corrected stochastic block model and compare several community detection criteria under both the standard and the degree-corrected models. We show which criteria are consistent under which models and constraints, as well as compare their relative performance in practice. We find that methods based on the degree-corrected block model, which includes the standard block model as a special case, are consistent under a wider class of models and that modularity-type methods require parameter constraints for consistency, whereas likelihood-based methods do not. On the other hand, in practice, the degree correction involves estimating many more parameters, and empirically we find it is only worth doing if the node degrees within communities are indeed highly variable. We illustrate the methods on simulated networks and on a network of political blogs.

Community extraction for social networks.

Analysis of networks and in particular discovering communities within networks has been a focus of recent work in several fields and has diverse applications. Most community detection methods focus on partitioning the entire network into communities, with the expectation of many ties within communities and few ties between. However, many networks contain nodes that do not fit in with any of the communities, and forcing every node into a community can distort results. Here we propose a new framework that extracts one community at a time, allowing for arbitrary structure in the remainder of the network, which can include weakly connected nodes. The main idea is that the strength of a community should depend on ties between its members and ties to the outside world, but not on ties between nonmembers. The proposed extraction criterion has a natural probabilistic interpretation in a wide class of models and performs well on simulated and real networks. For the case of the block model, we establish asymptotic consistency of estimated node labels and propose a hypothesis test for determining the number of communities.