Learning Changes in Graphon Attachment Network Models
Authors: Xinyuan Fan, Bufan Li, Chenlei Leng, Weichi Wu
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments demonstrate the effectiveness of our approach across various network settings, highlighting its potential for dynamic network analysis. Numerical experiments demonstrate the effectiveness of our approach across various network settings, highlighting its potential for dynamic network analysis. Section 4 presents the findings from our simulation studies. Section 5 A real data example. |
| Researcher Affiliation | Academia | 1Department of Statistics and Data Science, Tsinghua University, Beijing, China 2Department of Statistics, University of Warwick, Coventry, UK. |
| Pseudocode | Yes | Detailed descriptions of these algorithms can be found in Algorithm 1 and Algorithm 2, provided in Appendix A.1. Algorithm 1 Random Interval Distillation for Piecewise Polynomial Signals. Algorithm 2 Localization Procedure. |
| Open Source Code | No | The paper does not explicitly provide concrete access to source code for the methodology described. |
| Open Datasets | Yes | The email-Eu-core network data (Paranjape et al., 2017) was constructed using email communications from a large European research institution. |
| Dataset Splits | No | The paper analyzes time-evolving networks in simulation studies and a real data example (email-Eu-core network) to detect change points. It does not use conventional training/test/validation splits; instead, it processes the entire evolving network to identify structural changes over time. |
| Hardware Specification | No | The paper does not provide specific hardware details used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers needed to replicate the experiment. |
| Experiment Setup | Yes | In each simulation setting, we repeat the experiment 100 times and report the mean values for H(ˆη, η)/T, ARI, and the difference ˆK K. For the subgraphs, we choose H = , , or , as these are commonly used in the literature (e.g., Maugis et al. (2020), Shao et al. (2022)). The threshold τ in Algorithm 1 is chosen based on a data-driven approach inspired by Fan & Wu (2024). Specifically, we use τ = max j=1,...,T h max j<t<j+h | Xt j,j+h(H)| log(T), where h = 3 log T or 6 log T . When we suspect severe degree heterogeneity, we use h = 6 log T ; otherwise, we use h = 3 log T . Due to space constraints, the specific settings for the four simulations are provided in Appendix A.2. Scenario 1: Model with Degree Heterogeneity. We set T = 400 and η = 300, and define the graphon function h T,t(x, y) as follows: h T,t(x, y) = ( 1 T 0.6(xy) α, t < 300, 1 T 0.6(xy) 0.9, t 300, (7) where α = 0.75, 0.65, 0.5. |