Alipour Fahimeh, Dokshin Fedor, Maleki Zeinab, Song Yunya, Ramazi Pouria
Department of Electrical and Computer Engineering, Isfahan University of Technology, Khomeyni Shahr, Daneshgah e Sanati Hwy, Isfahan PG9G+39R, Isfahan, Iran.
Department of Sociology, University of Toronto, 700 University Ave, Toronto, Ontario, Canada M5G 1Z5.
PNAS Nexus. 2024 Oct 22;3(10):pgae428. doi: 10.1093/pnasnexus/pgae428. eCollection 2024 Oct.
Behavioral diffusion is commonly modeled with the linear threshold model, which assumes that individuals adopt a behavior when enough of their social contacts do so. We observe, however, that in many common empirical settings individuals also appear to abandon a behavior when too many of their close contacts exhibit it. The bi-threshold model captures this tendency by adding an upper threshold, which, when exceeded, triggers behavioral disadoption. Here we report an empirical test of the bi-threshold model. We overcome the significant challenge of estimating individuals' heterogeneous thresholds by extending a recently introduced decision-tree based algorithm to the bi-threshold setting. Using the context of the spread of news about three different topics on social media (the Higgs boson, the Melbourne Cup horse race, and the COVID-19 vaccination campaign in China), we show that the bi-threshold model predicts user engagement with the news orders of magnitude more accurately than the linear threshold model. We show that the performance gains are due especially to the bi-threshold model's comparative advantage in predicting behavioral decline, an important but previously overlooked stage of the behavioral diffusion cycle. Overall, the results confirm the existence of the second upper threshold in some contexts of diffusion of information and suggest that a similar mechanism may operate in other decision-making contexts.
行为扩散通常用线性阈值模型来建模，该模型假设当个体的足够多的社会联系人采取某种行为时，个体也会采取该行为。然而，我们观察到，在许多常见的实证环境中，当个体的太多亲密联系人表现出某种行为时，个体似乎也会放弃该行为。双阈值模型通过添加一个上限阈值来捕捉这种趋势，当超过这个上限阈值时，就会触发行为摒弃。在此，我们报告双阈值模型的实证检验。我们通过将最近引入的基于决策树的算法扩展到双阈值设置，克服了估计个体异质阈值的重大挑战。利用社交媒体上关于三个不同主题（希格斯玻色子、墨尔本杯赛马和中国的新冠疫苗接种运动）的新闻传播背景，我们表明双阈值模型预测用户对新闻的参与度比线性阈值模型精确几个数量级。我们表明，性能提升尤其归功于双阈值模型在预测行为下降方面的比较优势，行为下降是行为扩散周期中一个重要但此前被忽视的阶段。总体而言，结果证实了在某些信息传播背景下存在第二个上限阈值，并表明类似机制可能在其他决策背景中起作用。
