Restoring: A Greedy Heuristic Approach Based on Neighborhood for Correlation Clustering

Abstract

Correlation Clustering has received considerable attention in machine learning literature due to its not requiring specifying the number of clusters in advance. Many approximation algorithms for Correlation Clustering have been proposed with worst-case theoretical guarantees, but with less experimental evaluations. These methods simply consider the direct associations between vertices and achieve poor performance in real datasets. In this paper, we propose a neighborhood-based method called Restoring, in which we argue that the neighborhood around two connected vertices is important and two vertices belonging to the same cluster should have the same neighborhood. Our algorithm iteratively chooses two connected vertices and restores their neighborhood. We also define the cost of keeping or removing one non-common neighbor and identify a restoring order based on the neighborhood similarity. Experiments conducted on five sub datasets of Cora show that our method performs better than existing well-known methods both in results quality and objective value.