Local Differential Privacy-Preserving Spectral Clustering for General Graphs

Authors: Sayan Mukherjee, Vorapong Suppakitpaisarn

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical tests further corroborate these theoretical findings.
Researcher Affiliation Academia Sayan Mukherjee EMAIL The University of Tokyo Vorapong Suppakitpaisarn EMAIL The University of Tokyo
Pseudocode Yes Compute the second smallest eigenvector v = [v1, . . . , vn] of LG using the Lanczos algorithm or an approximation method, such as the one proposed in Adil & Saranurak (2024). Let v1, . . . , vn be distinct nodes of V such that vv1 vvn. Return the cut (S, S), where S = {v1, . . . , i0} and i0 = arg min 1 i n αG(v1, . . . , vi).
Open Source Code No The paper does not contain any explicit statements or links indicating that open-source code for the described methodology is provided.
Open Datasets Yes We conduct experiments on real social networks to verify our theoretical results. In this work, we mainly use the network called Social circles: Facebook obtained from the Stanford network analysis project (SNAP), detailed in Leskovec & Mcauley (2012).
Dataset Splits Yes For each probability p and graph, we create 100 random graphs F with the given probability.
Hardware Specification No The paper does not contain any specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes We examine p {0.0001q : 1 q 50}. For each probability p and graph, we create 100 random graphs F with the given probability. Note that the original graph is represented by G. We then compute the difference between the clustering results of G (represented by SC2(G)) and that of G F (represented by SC2(G F)).