Abstract:We revisit the simultaneous approximation model for the correlation clustering problem introduced by Davies, Moseley, and Newman[DMN24]. The objective is to find a clustering that minimizes given norms of the disagreement vector over all vertices. We present an efficient algorithm that produces a clustering that is simultaneously a $63.3$-approximation for all monotone symmetric norms. This significantly improves upon the previous approximation ratio of $6348$ due to Davies, Moseley, and Newman[DMN24], which works only for $\ell_p$-norms. To achieve this result, we first reduce the problem to approximating all top-$k$ norms simultaneously, using the connection between monotone symmetric norms and top-$k$ norms established by Chakrabarty and Swamy [CS19]. Then we develop a novel procedure that constructs a $12.66$-approximate fractional clustering for all top-$k$ norms. Our $63.3$-approximation ratio is obtained by combining this with the $5$-approximate rounding algorithm by Kalhan, Makarychev, and Zhou[KMZ19]. We then demonstrate that with a loss of $\epsilon$ in the approximation ratio, the algorithm can be adapted to run in nearly linear time and in the MPC (massively parallel computation) model with poly-logarithmic number of rounds. By allowing a further trade-off in the approximation ratio to $(359+\epsilon)$, the number of MPC rounds can be reduced to a constant.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the **simultaneous approximation of multiple norms in the correlation clustering problem**. Specifically, the authors hope to find a clustering method that can minimize the disagreement vector under a given norm, and this clustering method can provide a good approximation for all monotone symmetric norms simultaneously. #### Relevant background - **Correlation clustering problem**: This is a classic unsupervised machine learning problem, with the goal of classifying data elements according to the similarity between them. For each pair of data points, we have a similarity label (+ or -), indicating whether they should be grouped into the same class. The goal of clustering is to minimize the number of inconsistent edges, that is, the situation where similar points are grouped into different classes or dissimilar points are grouped into the same class. - **Choice of norms**: The traditional correlation clustering problem usually uses the $\ell_1$ norm to measure the cost of the disagreement vector. However, recent research has begun to consider other norms, such as the $\ell_p$ norm ($p\in[1,\infty]$), and even more general monotone symmetric norms. #### Main contributions of the paper 1. **Expanded the range of norms**: The authors not only considered the $\ell_p$ norm, but also extended to all monotone symmetric norms. This makes their method more general and powerful. 2. **Improved the approximation ratio**: They proposed an efficient algorithm that can construct a clustering with a 63.3 - fold approximation for all monotone symmetric norms in polynomial time. Compared with previous work (for example, Davies, Moseley, and Newman [DMN24]), this approximation ratio is significantly reduced. 3. **Implementation in the parallel computing model**: The authors further demonstrated how to implement this algorithm in the Massively Parallel Computation (MPC) model, enabling it to run in near - linear time and requiring only a constant number of communication rounds. Moreover, by sacrificing a certain approximation ratio, the number of rounds can also be reduced to a constant level. #### Markdown - format display of formulas The formulas involved in the paper are as follows: - **Disagreement vector**: \[ \text{cost}_C(u)=\sum_{uv\in E^+}\mathbf{1}(u\text{ and }v\text{ belong to different classes})+\sum_{uv\in E^-}\mathbf{1}(u\text{ and }v\text{ belong to the same class}) \] where $E^+$ represents the set of similar edges, and $E^-$ represents the set of dissimilar edges. - **top - k norm**: \[ \text{cost}^k_x = \max_{T\subseteq V,|T| = k}\sum_{u\in T}\text{cost}_x(u) \] - **Triangle inequality constraint**: \[ x_{uv}+x_{uw}\geq x_{vw},\quad\forall u,v,w\in V \] These formulas ensure the approximation performance of the clustering results under different norms and provide a theoretical basis for the subsequent algorithm design.

Simultaneously Approximating All Norms for Massively Parallel Correlation Clustering

Simultaneously Approximating All $\ell_p$-norms in Correlation Clustering

Combinatorial Correlation Clustering

Massively Parallel Correlation Clustering in Bounded Arboricity Graphs

Correlation Clustering with Sherali-Adams

Handling Correlated Rounding Error Via Preclustering: A 1.73-Approximation for Correlation Clustering

Multilayer Correlation Clustering

Fully Dynamic Correlation Clustering: Breaking 3-Approximation

Fast Combinatorial Algorithms for Min Max Correlation Clustering

Parameterized Approximation Schemes for Clustering with General Norm Objectives

Clustering to Minimize Cluster-Aware Norm Objectives

Improved Approximation Algorithms for Bipartite Correlation Clustering

A 4-approximation algorithm for min max correlation clustering

Constant Approximation for Normalized Modularity and Associations Clustering

Parameterized Correlation Clustering in Hypergraphs and Bipartite Graphs

Understanding the Cluster Linear Program for Correlation Clustering

A Scalable Approach for General Correlation Clustering

Parameterized Approximation for Robust Clustering in Discrete Geometric Spaces

Dynamic Correlation Clustering in Sublinear Update Time

Understanding the Cluster LP for Correlation Clustering

A Combinatorial Algorithm for Weighted Correlation Clustering