ForceAtlas2: A Continuous Graph Layout Algorithm for Gephi

ForceAtlas2 is a continuous, force-directed graph layout algorithm developed for Gephi. Unlike algorithms designed primarily for static performance, ForceAtlas2 prioritizes high-quality, continuous spatialization to support interactive exploration and trial-and-error analysis. It integrates established techniques such as Barnes Hut simulation, degree-dependent repulsion, and adaptive temperature control to provide a user-friendly experience for networks containing between 10 and 10,000 nodes.

The algorithm treats nodes as charged particles that repel each other and edges as springs that provide attraction. This physical simulation converges to a state that visually highlights structural communities. It employs a (1, -1) energy model, placing it between the Fruchterman-Reingold and LinLog models, effectively balancing visual cluster density with overall network structure.

ForceAtlas2 includes several user-configurable settings: LinLog mode for better clustering, gravity to pull peripheral nodes toward the center, edge weight influence, and a 'Dissuade Hubs' setting to shift hubs to the periphery. It also provides advanced features such as preventing node overlap and approximating repulsion through the Barnes Hut optimization to ensure smooth performance on larger datasets.

The algorithm addresses the trade-off between speed and precision through adaptive step lengths. By measuring node 'swinging' (the divergence between consecutive force vectors), the algorithm dynamically adjusts local and global speeds. This prevents oscillations and allows the layout to remain responsive and fluid while it converges, making it highly suitable for Gephi's real-time interaction requirements.

Benchmarking against Fruchterman-Reingold and Yifan Hu algorithms demonstrates that ForceAtlas2 provides an excellent compromise between layout quality and computational speed. While Yifan Hu is slightly faster for small networks, ForceAtlas2 offers superior visual quality and robustness across various network scales, remaining effective even up to 100,000 nodes when given sufficient computation time.