Figure 1.
Layouts with different types of forces.
Layouts with Fruchterman-Reingold (), ForceAtlas2 (
) and the LinLog mode of ForceAtlas2 (
).
Figure 2.
Regular repulsion vs. repulsion by degree.
Fruchterman-Rheingold layout on the left (regular repulsion) and ForceAtlas2 on the right (repulsion by degree). While the global scheme remains, poorly connected nodes are closer to highly connected nodes. ().
Figure 3.
ForceAtlas2 with gravity at 2 and 5. Gravity brings disconnected components closer to the center (and slightly affects the shape of the components as a side-effect).
Figure 4.
ForceAtlas2 with scaling at 1, 2 and 10. The whole graph expands as scaling affects the distance between components as well as their size. Note that the size of the nodes remains the same; scaling is not zooming.
Figure 5.
Effects of the edge weight influence.
ForceAtlas2 with Edge Weight Influence at 0, 1 and 2 on a graph with weighted edges. It has a strong impact on the shape of the network.
Figure 6.
Effects of the overlapping prevention.
ForceAtlas2 without and with the nodes overlapping prevention.
Figure 7.
The oscillation of nodes increases with speed.
Fruchterman-Rheingold layout at speeds 100, 500 and 2,500 (superposition at two successive steps).
Figure 8.
Adaptive local speed is a good compromise.
Evolution of the quality of ForceAtlas2 variants at each iteration (the higher the better). Different values of the local speed give different behaviors. The adaptive local speed achieves the best compromise between performance and quality. The network used is “facebook_ego_0” from our dataset.
Figure 9.
Effects of adaptive local speed on different networks.
Evolution of the quality of ForceAtlas2 variants at each iteration on the other facebook ego-networks of our dataset. The adaptive local speed is always the best. Local speed 0.001 converges poorly because the speed is too low. Local speed 0.1 converges poorly because it oscillates a lot: the speed is too high. Local speed 0.01 is sometimes adapted to the network, and sometimes not, but never outperforms the adaptive speed.
Figure 10.
Evolution of the layout quality for a single network over 2048 steps. Rows are the 4 different layouts and columns the 3 different randomizations. The red dot is the “Quick and dirty point” where 50% of the maximum quality is reached, and the blue dot is the “Quasi-optimal point” where 90% of the maximum quality is reached. The full visualization is available at this URL: https://github.com/medialab/benchmarkForceAtlas2/tree/master/benchmarkResults.
Figure 11.
Overall results of the benchmark.
Note that the second and third charts have logarithmic scales. FR is really slow, YH has a good performance and FA2 has a good quality.
Figure 12.
Quasi-Optimal Time over network size.
The lower is the better. Note that both scales are logarithmic. On small networks, FR is the best while FA2_LL is slower. On large networks, FR has a poor performance while other algorithms perform similarly on large networks.
Figure 13.
Layouts give visibly different results.
We find that FA2_LL and FA2 are more readable, because the different areas of the network are more precisely defined. However, we do not know any quality measure that captures this phenomenon.