Spectral embedding for dynamic networks with stability guarantees

Authors: Ian Gallagher, Andrew Jones, Patrick Rubin-Delanchy

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Section 2 gives a pedagogical example demonstrating the cross-sectional and longitudinal stability of UASE in a two-step dynamic stochastic block model, while highlighting the instability of omnibus and independent spectral embedding. In Section 3, we prove a central limit theorem for UASE under a dynamic latent position model, and demonstrate that the distribution satisfies both stability conditions. In Section 4, we review the stability of other dynamic network embedding procedures. Section 5 presents an example of UASE applied to a dynamic network of social interactions in a French primary school, with a dynamic community detection example given in the main text and a further classification example provided in the Appendix.
Researcher Affiliation Academia Ian Gallagher University of Bristol, UK ian.gallagher@bristol.ac.uk Andrew Jones University of Bristol, UK andrew.jones@bristol.ac.uk Patrick Rubin-Delanchy University of Bristol, UK patrick.rubin-delanchy@bristol.ac.uk
Pseudocode Yes Algorithm 1 Unfolded adjacency spectral embedding for dynamic networks input symmetric adjacency matrices A(1), . . . , A(T ) 2 {0, 1}n n, embedding dimension d 1: Form the matrix A = (A(1)| |A(T )) 2 {0, 1}n T n (column concatenation) 2: Compute the truncated singular value decomposition UA AV> A of A, where A contains the d largest singular values of A, and UA, VA the corresponding left and right singular vectors 3: Compute the right embedding ˆY = VA 1/2 A 2 RT n d and create sub-embeddings ˆY(t) 2 Rn d where ˆY = ( ˆY(1); . . . ; ˆY(T )) (row concatenation) return node embeddings for each time period ˆY(1), . . . , ˆY(T )
Open Source Code No The paper does not explicitly state that the source code for the described methodology is available, nor does it provide a link to a code repository.
Open Datasets Yes The Lyon primary school data set shows the social interactions at a French primary school over two days in October 2009 [44]. ... The data are available for download from the Network Repository website1 [34]. 1https://networkrepository.com
Dataset Splits No The paper mentions simulating a dynamic network and using the Lyon primary school data set, but it does not specify any explicit training, validation, or test dataset splits for reproduction.
Hardware Specification Yes taking approximately five seconds on a 2017 Mac Book Pro
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., 'Python 3.8, PyTorch 1.9').
Experiment Setup Yes Given the unfolded adjacency matrix A = (A(1)| |A(20)), an estimated embedding dimension ˆd = 10 was obtained using profile likelihood [55] and we construct the embeddings ˆY(1), . . . , ˆY(20) 2 Rn 10, taking approximately five seconds on a 2017 Mac Book Pro.