Visual evaluation of graph representation learning based on the presentation of community structures

Abstract

Various graph representation learning models convert graph nodes into vectors using techniques like matrix factorization, random walk, and deep learning. However, choosing the right method for different tasks can be challenging. Communities within networks help reveal underlying structures and correlations. Investigating how different models preserve community properties is crucial for identifying the best graph representation for data analysis. This paper defines indicators to explore the perceptual quality of community properties in representation learning spaces, including the consistency of community structure, node distribution within and between communities, and central node distribution. A visualization system presents these indicators, allowing users to evaluate models based on community structures. Case studies demonstrate the effectiveness of the indicators for the visual evaluation of graph representation learning models.