Online toxicity spreads through social interactions in ways that resemble contagion processes observed in epidemiology. Prior studies have applied compartmental epidemic models to analyze the propagation of toxic behavior in digital environments, yet these approaches typically assume that recovered individuals must return to a susceptible state before becoming toxic again. In practice, many users rapidly re-engage in toxic activity after temporary recovery. To capture this behavior, we extend the SEIRS epidemiological framework by introducing an immediate relapse mechanism that allows transitions directly from the recovered state to the infected state. Using four datasets spanning political protests, public health debates, and controversial online discussions, we compare the standard SEIRS model with the proposed SEIRS-relapse formulation. Results show that incorporating relapse dynamics reduces model error and improves robustness across different sets of parameter initializations. These findings suggest that relapse mechanisms provide a useful extension for modeling recurring toxicity patterns in online discourse.

PDF