Interevent time distributions of human multi-level activity in a virtual world

Studying human behaviour in virtual environments provides extraordinary opportunities for a quantitative analysis of social phenomena with levels of accuracy that approach those of the natural sciences. In this paper we use records of player activities in the massive multiplayer online game Pardus over 1,238 consecutive days, and analyze dynamical features of sequences of actions of players. We build on previous work were temporal structures of human actions of the same type were quantified, and extend provide an empirical understanding of human actions of different types. This study of multi-level human activity can be seen as a dynamic counterpart of static multiplex network analysis. We show that the interevent time distributions of actions in the Pardus universe follow highly non-trivial distribution functions, from which we extract action-type specific characteristic “decay constants”. We discuss characteristic features of interevent time distributions, including periodic patterns on different time scales, bursty dynamics, and various functional forms on different time scales. We comment on gender differences of players in emotional actions, and find that while male and female act similarly when performing some positive actions, females are slightly faster for negative actions. We also observe effects on the age of players: more experienced players are generally faster in making decisions about engaging and terminating in enmity and friendship, respectively.

💡 Research Summary

This paper investigates the temporal dynamics of human multi‑level activity by exploiting a unique, large‑scale data set from the massive multiplayer online game Pardus. Over a period of 1,238 consecutive days, the authors collected timestamped logs of seven distinct action types performed by 7,818 active players in the Artemis universe: private communication (C), attacks (A), trades (T), friend additions (F), enemy additions (E), friend deletions (D), and enemy deletions (X). Each action is associated with a cost in “action points” (AP) that are regenerated every six minutes, introducing a natural coupling between a player’s online presence and the timing of certain actions.

The authors first filter out low‑activity accounts (fewer than 50 actions) to focus on statistically robust behavior. In total, 8,373,209 actions were recorded, with communication accounting for roughly 80 % of all events. Inter‑event times τ (the interval between two consecutive actions of the same player, regardless of type) are computed with a one‑second resolution, yielding a range from 0 s to over 63 million seconds (≈2 years).

Analysis of the τ distribution reveals a heavy‑tailed, power‑law‑like form that changes slope depending on the time scale. When τ is binned in 6‑hour windows, the distribution follows a power law with exponent ≈ −2.09; when the focus is narrowed to the first 24 hours and binned in 1‑minute intervals, the exponent softens to ≈ −1.12. These two regimes reflect the influence of circadian and weekly cycles, which are clearly visible as periodic peaks in the inset of the 6‑hour plot. A pronounced local minimum around τ ≈ 7 hours corresponds to the typical “work‑day” rhythm: players tend to be offline during standard work hours and return in the evening, making 7‑hour gaps less common than 8‑10‑hour gaps.

To quantify burstiness, the authors compute the burstiness coefficient B = (σ − m)/(σ + m), where m and σ are the mean and standard deviation of τ, respectively. Across the population, B values cluster around 0.5–0.9, indicating a high degree of bursty dynamics far from the Poisson benchmark (B ≈ 0). Action‑type specific B values show that negative actions (attacks, enemy additions, enemy removals) are more bursty (B ≈ 0.6–0.7) than positive actions (communication, trade, friend addition) which display lower burstiness (B ≈ 0.4–0.5). This suggests that hostile interactions tend to occur in rapid clusters, whereas cooperative or economic activities are more evenly spaced.

The authors further fit exponential tails to the τ distributions for each action type, extracting decay constants λ (P(τ) ∝ e^{−λτ}). Larger λ indicates faster successive actions. Attacks exhibit λ ≈ 1.2 × 10⁻⁴ s⁻¹, while communication shows λ ≈ 3.5 × 10⁻⁵ s⁻¹, confirming that aggressive actions are executed on a shorter timescale than routine communication. These differences are attributed to game mechanics (e.g., AP costs, UI design) and the intrinsic urgency of hostile behavior.

Gender analysis reveals that males and females behave similarly for positive actions, but females execute negative actions (A, E, D) with slightly shorter τ (≈ 8 % faster on average), indicating a modest gender‑related speed advantage in hostile contexts. Age (measured as days since registration) also matters: veteran players (longer in‑game tenure) tend to resolve friendships and enmities more quickly, with τ reduced by 12–15 % compared to newcomers. This points to experience‑driven efficiency in managing social ties.

At the community level, the aggregate τ distribution shifts dramatically during in‑game wars. During these conflict periods, the frequency of attacks and enemy additions spikes, τ values compress, and overall burstiness rises, reflecting the impact of exogenous events on collective temporal patterns.

In sum, the study demonstrates that human multi‑level activity in a virtual world cannot be captured by a single, homogeneous Poisson process. Instead, it exhibits a rich tapestry of temporal structures shaped by action‑type specific priorities, resource constraints (AP regeneration), circadian/weekly rhythms, gender and experience effects, and large‑scale events such as wars. The authors argue that such virtual‑world data provide a non‑intrusive, high‑resolution laboratory for testing theories of human dynamics and multiplex network evolution, bridging the gap between social science and natural‑science methodologies. Future work is suggested to explore causal inter‑action dependencies (e.g., whether an attack triggers a friend addition) and to integrate these empirical findings into multiplex network models that can more faithfully reproduce the observed bursty, multi‑scale temporal patterns.