Layouts

• Use more iterations. The number of iterations is a parameter of the method minimizeEnergy of the classes MinimizerBarnesHut and MinimizerClassic (called in the main method of class LinLogLayout). While 100 iterations suffice for many graphs up to a few hundred nodes, larger or "difficult" graphs may require more iterations. Increasing the number of iterations is particularly helpful if the energy values printed by LinLogLayout still decrease notably during the last 5 percent of the iterations (e.g., from iteration 95 to iteration 100).
• Reduce or eliminate gravitation. The gravitation factor is a parameter of the constructor of the classes MinimizerBarnesHut and MinimizerClassic (called in the main method of class LinLogLayout). Gravitation attracts all nodes to the barycenter of the layout. This prevents nodes from being moved to far from the remaining graph, but also distorts the layout. To reduce this distortion, set the gravitation factor to a very small value like 0.01 or 0.001. To eliminate the distortion entirely, set the gravitation factor to 0.0 -- but only if the graph is connected, otherwise the unconnected components will fly away into infinity (which will likely cause the minimizer to produce warnings like "The node distances in the layout are extremely nonuniform.").
• Compute several layouts and choose the best. Because the layouts are initialized randomly, each run of LinLogLayout produces a different layout. Run LinLogLayout several times and choose the layout that has the best (smallest) energy after the final iteration. (LinLogLayout prints the current energy to the console at each iteration.)

Clusterings

• Compute several clusterings and choose the best. Unlike the layouter, the clusterer of LinLogLayout is deterministic by default, i.e. produces the same clustering at each run (on the same graph). However, it can be randomized by commenting in a line marked with "randomize contraction" in the class OptimizerModularity. After commenting in this line, run LinLogLayout several times and choose the clustering with the best (largest) modularity.