Modeling and Analyzing Collective Attention in Online Chats

Collective attention refers to the shared focus of multiple individuals on a common stimulus, conversation, or topic. It shapes how information circulates, interactions unfold, and decisions emerge within groups. In online environments, collective attention can appear as either focused, when participants engage intensively with a single topic, or unfocused, when attention is distributed across multiple discussions or themes. While collective attention has been widely studied in social networks and controlled experiments, its dynamics in online chats, such as WhatsApp, Telegram, and YouTube Live, remain underexplored. This work addresses this gap by mapping three foundational components of attention theory, alertness, processing capacity, and selectivity, onto observable features of online chats: chat activity, parallel conversations, and conversation topics. Using this mapping, the work defines and identifies conversation periods, intervals of sustained, synchronous message exchange that serve as analytically meaningful units for studying attention. From this foundation, the work offers two primary contributions. First, it introduces a computational model to classify each conversation period as either focused or unfocused, enabling automated detection of attention dynamics across chat platforms. Second, it characterizes how the three components of attention emerge across four case studies: two from WhatsApp and Telegram group chats (covering political and non-political discussions), and two from YouTube Live chats, including a Brazilian COVID-19 parliamentary inquiry (CPI) and FIFA 2025 World Club Cup matches. The classifiers for alertness, processing capacity, and selectivity achieve AUC scores ranging from 0.76 to 0.84. Temporal activity patterns (alertness) and conversational thread coherence (processing capacity) provide the strongest, closely comparable signals, whereas thematic consistency across a conversation period (selectivity) proves comparatively harder to capture from text alone. A cross-platform comparison reveals that platform structure strongly shapes both the prevalence and character of collective attention. YouTube Live chats, anchored to externally paced live events, exhibit substantially higher rates of focused periods, reflecting short-lived but intense bursts of engagement synchronized with the unfolding event. WhatsApp and Telegram groups, by contrast, are not bound to an external synchronizing stimulus and exhibit markedly lower rates of focused periods, particularly in politically oriented discussions, where conversations are more prone to thematic drift and structural fragmentation across a broader, less constrained range of topics. These differences demonstrate that collective attention is platform-contingent: it adopts different temporal and structural configurations depending on how each environment mediates visibility, synchrony, and thematic scope. Beyond theoretical contributions, the work offers computational tools for analyzing synchronous group communication. Understanding how attention becomes focused or fragmented provides a basis for studying information diffusion, group coordination, and political polarization. The models and methods developed here have broader applications in areas such as digital discourse analysis, advertising, and the detection of misinformation and harmful content in live conversational settings.