Core Concepts
Efficiently recover clusters in well-clustered graphs with a differentially private algorithm.
Abstract
The content discusses a differentially private clustering algorithm tailored for well-clustered graphs. It explores the theoretical background of spectral clustering and differential privacy, presenting an algorithm that achieves privacy guarantees without compromising utility. Experimental evaluations validate the algorithm's effectiveness in recovering clusters accurately while maintaining privacy.
Introduction
Graph clustering is fundamental in unsupervised machine learning.
Various measures evaluate the quality of graph clusterings.
Well-Clustered Graphs
Definition of well-clustered graphs.
Previous works on spectral clustering for such graphs.
Differential Privacy
Introduction to differential privacy and its importance.
Theoretical foundation for applying DP to graph clustering.
Algorithm Description
SDP formulation for extracting cluster structure.
Incorporating noise to ensure privacy.
Proof of Theorem 1
Privacy analysis of the algorithm.
Utility analysis based on approximation ratio and trade-offs.
Experiments
Evaluation of the algorithm's performance on synthetic datasets.
Stats
Our private mechanism almost matches the best-known non-private one in terms of approximate accuracy or utility.
Quotes
"Our private mechanism is inspired by the recent work of Chen et al."
"Any (pure) ε-DP algorithm entails substantial error in its output."