Recent Progress on Correlation Clustering: From Local Algorithms to Better Approximation Algorithms and Back (Invited Talk)

Abstract

Correlation clustering is a classic model for clustering problems arising in machine learning and data mining. Given a set of data elements represented as vertices of a graph and pairwise similarity represented as edges, the goal is to find a partition of the vertex set so as to minimize the total number of edges across the parts plus the total number of non-edges within the parts.
Introduced in the early 2000s [Bansal et al., 2004], correlation clustering has received a large amount of attention through the years. A natural linear programming relaxation was shown to have an integrality gap of at least 2 and at most 2.5 [Ailon et al., 2008] in 2005, and in 2015 at most 2.06 [Chawla et al., 2015].
In 2021, motivated by large-scale application new structural insights allowed to derive a simple, practical algorithm that achieved an O(1)-approximation in a variety of models (Massively Parallel, Sublinear, Streaming or Differentially-private) [Vincent Cohen{-}Addad et al., 2021; Cohen-Addad et al., 2022]. These new insights turned out to be a key building block in designing better algorithms: It serves as a pre-clustering of the input graph that enables algorithm with approximation guarantees significantly better than 2 [Vincent Cohen{-}Addad et al., 2023; Vincent Cohen{-}Addad et al., 2022]. It is a key component in the new algorithm that achieves a 1.44-approximation [Nairen Cao et al., 2024] and in the new local-search based 1.84-approximation for the Massively Parallel, Sublinear, and Streaming models [Vincent Cohen{-}Addad et al., 2024]. This talk will review the above recent development and what are the main open research directions.
A collection of joint works with Nairen Cao, Silvio Lattanzi, Euiwoong Lee, Shi Li, David Rasmussen Lolck, Slobodan Mitrovic, Alantha Newman, Ashkan Norouzi-Fard, Nikos Parotsidis, Marcin Pilipczuk, Jakub Tarnawski, Mikkel Thorup, Lukas Vogl, Shuyi Yan, Hanwen Zhang.