Figure 5.

(a) Decision tree of the dynamics of the proposed agent-based Heider Balance model. Our model algorithm processes the network iteratively in update steps. First, a random triad ($▴$) is selected. It can either be balanced () or unbalanced (). If the triad is unbalanced, we select its edge (drawn in black in the figure). For the triad with two positive edges, the chosen edge is positive or negative with probabilities $(1-p_n)$ or $p_n$, respectively. For the triad with three negative edges, we choose one of them randomly. Then, we either remove the selected edge with probability $p_r$ or change an opinion of one of the edge endpoints with probability $(1-p_r)$ so that changing the edge’s sign will be possible. For a full description, see the main text. At the end of the single update with probability $p_{\mathit{add}}$, a new edge is added to the graph. (b) Transition probabilities in the NetSense data lie between the results of the agent-based model with HBT dynamics and of the random models. Plots show transition probabilities from unbalanced $T(u\to b)$ or balanced $T(b\to b)$) triads to balanced ones as a function of the tolerance $\mathrm{\Theta }$. The solid blue lines with circles represent empirical results. Dashed red and dotted green lines show results for node- and edge-randomized models, respectively. Solid purple lines with diamonds show the results of our agent-based model (based on diagram 5a). Fuzzy markers are drawn where too few triads of a given type make the obtained probabilities unreliable. Results imply that the dynamics of triads recorded in the empirical data set are driven by both Heider interactions and processes generating randomized data. Each model data point was obtained by at least 100 simulation realizations. The shaded area shows standard deviations of ABM transition probabilities.