Nonconvex Sparse Graph Learning under Laplacian Constrained Graphical Model

Authors: Jiaxi Ying, José Vinícius de Miranda Cardoso , Daniel Palomar

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments involving synthetic and realworld data sets demonstrate the effectiveness of the proposed method.
Researcher Affiliation Academia Jiaxi Ying1 jx.ying@connect.ust.hk José Vinícius de M. Cardoso1 jvdmc@connect.ust.hk Daniel P. Palomar1,2 palomar@ust.hk Department of Electronic and Computer Engineering1 Department of Industrial Engineering and Decision Analytics2 The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong
Pseudocode Yes Algorithm 1 Nonconvex Graph Learning (NGL)
Open Source Code Yes An open source R package is available at https://github.com/mirca/sparse Graph.
Open Datasets Yes We conduct numerical experiments on the 2019-n Co V data set1 from 98 anonymous Chinese patients affected by the outbreak of 2019-n Co V on early February, 2020. [...] 2019-n Co V data is available in a queryable format via the R package n Cov2019 which lives on Git Hub: https://github.com/Guangchuang Yu/n Cov2019.
Dataset Splits No The paper generates synthetic data or uses a real-world dataset to form a sample covariance matrix (S) for learning the graph. It does not explicitly mention splitting the data into distinct training, validation, and test sets for model development and evaluation.
Hardware Specification Yes The computational time for GLE-ADMM, NGL-SCAD and NGL-MCP are 2.9, 0.7 and 0.8 seconds, respectively, conducted on a PC with a 2.8 GHz Inter Core i7 CPU and 16 GB RAM.
Software Dependencies No The paper mentions 'An open source R package' and 'R package n Cov2019' but does not specify version numbers for R itself, these packages, or any other software dependencies.
Experiment Setup Yes We set γ equal to 1.01 in h MCP,λ(x) and 2.01 in h SCAD,λ(x) for all the experiments. [...] The number of nodes in the Barabasi-Albert graph is p = 50, and the weights associated with edges are uniformly sampled from U(2, 5). The sample covariance matrix is constructed by S = 1/n XX, where n is the number of samples. [...] The sample size ratio is n/p = 100. [...] The regularization parameter λ for each algorithm is fine-tuned.