Change Point Detection on A Separable Model for Dynamic Networks
Authors: Yik Lun Kei, Hangjian Li, Yanzhen Chen, OSCAR HERNAN MADRID PADILLA
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on both simulated and real data show good performance of the proposed framework, and an R package CPDstergm is developed to implement the method. |
| Researcher Affiliation | Collaboration | Yik Lun Kei EMAIL Department of Statistics University of California, Santa Cruz Hangjian Li EMAIL Walmart Global Tech Yanzhen Chen EMAIL Department of Information Systems, Business Statistics and Operations Management Hong Kong University of Science and Technology Oscar Hernan Madrid Padilla EMAIL Department of Statistics and Data Science University of California, Los Angeles |
| Pseudocode | Yes | The algorithm to solve (6) via ADMM is presented in Algorithm 1. Algorithm 1 Group Fused Lasso STERGM 1: Input: initialized parameters θ(1), γ(1), β(1), u(1), tuning parameter λ, penalty parameter α, number of iterations for ADMM, Newton-Raphson, and Group Lasso A, C, D, vectorized network data # y, network change statistics H 2: for a = 1, , A do 3: # θ = vecτp(θ(a)), # z (a) = vecτp(1τ,1γ(a) + Xβ(a)), # u(a) = vecτp(u(a)) 4: for c = 1, , C do 5: Let # θ c+1 be updated according to (13) 6: end for 7: θ(a+1) = vec 1 τ,p(# θ c+1) 8: Set γ = γ(a) and β = β(a) 9: for d = 1, , D do 10: Let βd+1 i, be updated according to (14) for i = 1, , τ 1 11: γd+1 = (1/τ)11,τ (θ(a+1) + u(a) X βd+1) 12: end for 13: γ(a+1) = γd+1, β(a+1) = βd+1 14: z(a+1) = 1τ,1γ(a+1) + Xβ(a+1) 15: u(a+1) = θ(a+1) z(a+1) + u(a) 16: end for 17: ˆθ θ(a+1) 18: Output: learned parameters ˆθ |
| Open Source Code | Yes | The R package CPDstergm is available at https://github.com/allenkei/CPDstergm. The code to reproduce the results in the manuscript using the R package is available at https://github.com/allenkei/CPDstergm_demo. |
| Open Datasets | Yes | The MIT cellphone data is available in the R package CPDstergm, and the stock market data is available in the R package ecp. |
| Dataset Splits | No | The paper does not provide explicit training/test/validation dataset splits in the traditional machine learning sense for its model. For simulated data, it defines the location of true change points (e.g., "K = 3 true change points are located at t = {26, 51, 76}") which implies partitions for evaluation (A1, A2, A3, A4), but this is not a train/test/validation split for model training. For real data, no split information is given. |
| Hardware Specification | No | The paper includes a table of 'Time comparison in seconds across different methods and node sizes' (Table 10), but it does not specify the hardware (CPU, GPU, etc.) on which these computations were performed. Therefore, no specific hardware details are provided for reproducibility. |
| Software Dependencies | Yes | The R library ergm (Handcock et al., 2022) provides an extensive list of network statistics that boost the power of the proposed method... R package tergm (Krivitsky & Handcock, 2022). (Citation [Handcock et al., 2022] mentions: "R package version 4.3.2." and [Krivitsky and Handcock, 2022] mentions: "R package version 4.1.0.") |
| Experiment Setup | Yes | Specifically, to detect change points with the proposed method, we initialize the penalty parameter α = 10, and we let the tuning parameter λ = 10b with b ∈ {0, 1, 2, 3, 4}. For each λ, we run A = 200 iterations of ADMM and the stopping criterion in (56) uses ϵtol = 10^−7. Within each ADMM iteration, we run C = 20 iterations of the Newton-Raphson method, and D = 20 iterations for the Group Lasso update. The stopping criteria for the Newton-Raphson method is ||# θ^(c+1) - # θ^c||_2 < 10^−3. To construct the data-driven threshold in (18), we use the 0.9 quantile of the standard Normal distribution. ... When the spacing between consecutive change points is less than a threshold or B̂_k - B̂_{k−1} < δ_spc, we keep the detected change point with greater ζ̂ value to avoid clusters of nearby change points. Furthermore, as the endpoints of a time span are usually not of interest, we discard the change point B̂_k that is less than a threshold δ_end and greater than T - δ_end. In Section 5, we set δ_spc = 5, and we set δ_end = 5 and δ_end = 10 for the simulated and real data experiments, respectively. |