QUIC: Quadratic Approximation for Sparse Inverse Covariance Estimation

Authors: Cho-Jui Hsieh, Mátyás A. Sustik, Inderjit S. Dhillon, Pradeep Ravikumar

JMLR 2014 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that our method is superlinearly convergent, and present experimental results using synthetic and real-world application data that demonstrate the considerable improvements in performance of our method when compared to previous methods. [...] We summarize the experimental results in Section 5, where we compare the algorithm using both real data and synthetic examples from Li and Toh (2010).
Researcher Affiliation Academia Cho-Jui Hsieh EMAIL Matyas A. Sustik EMAIL Inderjit S. Dhillon EMAIL Pradeep Ravikumar EMAIL Department of Computer Sciences University of Texas at Austin Austin, TX 78712, USA
Pseudocode Yes Algorithm 1: QUadratic approximation for sparse Inverse Covariance estimation (QUIC overview) [...] Algorithm 2: QUadratic approximation for sparse Inverse Covariance estimation (QUIC) [...] Algorithm 3: General Block Quadratic Approximation method for Sparse Inverse Covariance Estimation
Open Source Code Yes Our software package QUIC with MATLAB and R interface1 is public available at http://www.cs.utexas.edu/~sustik/QUIC/. The QUIC R package is also available from CRAN.
Open Datasets Yes We use the real world biology data sets preprocessed by Li and Toh (2010) to compare the performance of our method with other state-of-the-art methods.
Dataset Splits No The paper describes generating synthetic data and using preprocessed real-world biology datasets, but it does not specify explicit training/test/validation splits, proportions, or methodologies for partitioning the data to reproduce experiments.
Hardware Specification Yes all experiments were executed on 2.83GHz Xeon X5440 machines with 32G RAM and Linux OS.
Software Dependencies Yes We have implemented QUIC in C++ with MATLAB interface, and all experiments were executed on 2.83GHz Xeon X5440 machines with 32G RAM and Linux OS. We use the latest version glasso 1.7 downloaded from http://www-stat.stanford.edu/~tibs/glasso/.
Experiment Setup Yes We set the augmented Lagrangian parameter to ρ = 50 and the over-relaxation parameter to α = 1.5. These parameters achieved the best speed on the ER data set. [...] We use the identity matrix as the initial point for QUIC, ADMM, SINCO, and IPM. [...] In the first set of experiments, we set the regularization parameter λ to be 0.5 [...] we set the number of coordinate descent steps to be αt for the t-th outer iteration, where α is a constant; we use α = 1/3 in our experiments.