Should I Stay or Should I Go: Predicting Changes in Cluster Membership

TL;DR

This paper introduces a novel approach to predict individual node movements within communities, focusing on whether nodes stay, move, or drop out, using local/global features and embeddings.

Contribution

It proposes new classification features, including embedding-based distances and historical chains, for predicting node cluster membership changes.

Findings

Embedding-based features improve prediction accuracy.

Node history chains capture movement patterns effectively.

Different problem formulations vary in complexity and performance.

Abstract

Most research on predicting community evolution focuses on changes in the states of communities. Instead, we focus on individual nodes and define the novel problem of predicting whether a specific node stays in the same cluster, moves to another cluster or drops out of the network. We explore variations of the problem and propose appropriate classification features based on local and global node measures. Motivated by the prevalence of machine learning approaches based on embeddings, we also introduce efficiently computed distance-based features using appropriate node embeddings. In addition, we consider chains of features to capture the history of the nodes. Our experimental results depict the complexity of the different formulations of the problem and the suitability of the selected features and chain lengths.