Characterizing engagement dynamics across topics on Facebook

Abstract

Social media platforms heavily changed how users consume and digest information and, thus, how the popularity of topics evolves. In this paper, we explore the interplay between the virality of controversial topics and how they may trigger heated discussions and eventually increase users' polarization. We perform a quantitative analysis on Facebook by collecting ∼57M posts from ∼2M pages and groups between 2018 and 2022, focusing on engaging topics involving scandals, tragedies, and social and political issues. Using logistic functions, we quantitatively assess the evolution of these topics finding similar patterns in their engagement dynamics. Finally, we show that initial burstiness may predict the rise of users' future adverse reactions regardless of the discussed topic.

Fig 1. Summary of the analysis workflow followed in the current study.

News articles are collected from the GDELT Database, and their corpus is extracted, cleaned and analyzed to retrieve the most representing terms. The bipartite projection of the co-occurrence network built upon these terms serves as an input for the Louvain community detection algorithm to identify keyword clusters. Independent labellers then analyze these clusters to identify the subset of words that represent the topic under consideration, which are then used on Crowdtangle to retrieve the Facebook posts relating to those events.

Fig 2. Representation of a sample of four topics employing their normalized cumulative evolution of engagements and fittings.

The incidence of the α parameter can be observed in the sharpness of the fitting curves. The β parameter instead regulates the shift of the function through the x axis: the higher its value, the higher the delay from t₀ where the sigmoid produces its increment.

Fig 3. Joint distribution of α and β parameters obtained from the NLS regression for each topic.

We observe that topics are generally characterized by values of α and β, which explains how user interest in a topic does not increase all of a sudden but is the result of a process that evolves over time.

Fig 4

Upper panel: correlation between SI and LH score for each identified topic. Lower panel: correlation between SI and LH score for the top 4 most frequent topics. Overall, we observe how users react negatively as topics become sharply viral.