Extended Data Fig. 2: Effects of the edge thresholds and number of recurrences on the graph network’s performance.
From: Unveiling the predictive power of static structure in glassy systems
Pearson correlation coefficient for the ‘none’ network (no features, top) and the default network (distance and type features, bottom) for the ballistic, glassy and diffusive timescales (from left to right). Each point shows the median value of 3-10 independently trained networks. While in the first case there are some sharp peaks upon varying the edge threshold, in the second case the prediction monotonously increases with the edge threshold. Note that the prediction’s quality always increases with the number of recurrences.
