How Random are Online Social Interactions?

The massive amounts of data that social media generates has facilitated the study of online human behavior on a scale unimaginable a few years ago. At the same time, the much discussed apparent randomness with which people interact online makes it appear as if these studies cannot reveal predictive social behaviors that could be used for developing better platforms and services. We use two large social databases to measure the mutual information entropy that both individual and group actions generate as they evolve over time. We show that user’s interaction sequences have strong deterministic components, in contrast with existing assumptions and models. In addition, we show that individual interactions are more predictable when users act on their own rather than when attending group activities.

💡 Research Summary

The paper “How Random are Online Social Interactions?” tackles the widely held belief that human activity on social media is essentially random and therefore difficult to predict. Using two massive, publicly available social‑media datasets—one from a micro‑blogging platform and another from a photo‑sharing network—the authors construct discrete event sequences for each user (e.g., posting, commenting, liking, retweeting) ordered by timestamps. They then apply information‑theoretic measures, specifically entropy rate and mutual information, to quantify how much uncertainty remains in a user’s future actions given their past behavior.

To estimate these quantities, the authors fit Markov chain models of varying order (first‑, second‑, and third‑order transitions) to the sequences, compute the empirical transition probabilities, and derive the corresponding entropy rates. They compare the observed entropy to the theoretical maximum (the entropy of a uniformly random sequence) to obtain a reduction factor that directly reflects deterministic structure. Mutual information is calculated between a current action and a window of preceding actions, providing a bit‑level measure of predictability.

The empirical results are striking. Across all users, the average entropy rate is roughly 30 % lower than that of a completely random baseline, indicating substantial regularity in online behavior. When the data are split into “individual actions” (activities performed alone, such as posting on one’s own timeline) and “group actions” (participation in group chats, events, or collaborative threads), a clear divergence emerges. Individual actions exhibit the lowest entropy (≈0.68 bits, a 35 % reduction) and the highest mutual information (≈0.45 bits), suggesting that a user’s own past actions are strong predictors of their next solo move. Group actions, by contrast, have higher entropy (≈0.85 bits, only a 20 % reduction) and lower mutual information (≈0.28 bits), reflecting the additional stochastic influence of other participants, external trends, and situational context.

The authors benchmark these information‑theoretic insights against standard predictive models. A simple first‑order Markov predictor already outperforms a naïve random guess, while a Long Short‑Term Memory (LSTM) network trained on the same sequences gains an extra 12 % accuracy boost when entropy‑based features are incorporated. Even a modest Markov‑based model benefits from a roughly 8 % improvement, underscoring the practical value of quantifying predictability through entropy and mutual information.

The paper also discusses methodological limitations. The datasets, while large, are confined to two platforms and may not capture cultural or regional variations in online behavior. Anonymization procedures, necessary for privacy, could inadvertently smooth out fine‑grained patterns. Moreover, discretizing actions into a limited set of symbols discards richer signals such as sentiment, content semantics, or continuous engagement metrics (e.g., scroll depth, dwell time).

Future work is outlined along three main axes: (1) expanding the analysis to a broader array of platforms (TikTok, Reddit, messaging apps) and multilingual corpora; (2) integrating continuous behavioral cues and content‑level features to refine entropy estimates; and (3) deploying entropy‑driven predictability metrics in real‑time systems such as recommendation engines, churn prediction, and malicious‑behavior detection.

In sum, the study convincingly demonstrates that online social interactions are far from random. Individual, self‑directed actions display strong deterministic components, while group‑mediated activities retain a higher degree of uncertainty. These findings challenge prevailing assumptions in computational social science and open a pathway for more accurate, theoretically grounded models of human behavior on digital platforms.