رفتار انسان در جمعیت های فلش در جستجوی وب

This paper focuses on human behavior in web surfing during the special stage called Flash Crowd (FC) period. Some statistical properties of human behavior are investigated. A moving approximate entropy (ApEn) method is provided to precisely locate the FC stage at first. Then the multiscale entropy (MSE) method is applied to study the difference of behaviors between the FC stage and the Normal stage. The lower entropy value may imply that collective behavior in the FC stage tends to be more consistent and follows a process with self-similarity. Further investigation by MSE and interval time distribution on the collective level and the individual level reveals that the origin of FC formation is not due to the increasing number of users, but more likely the change of dynamic mechanism of individual behavior.

Revealing human behavior dynamics from big data of human activity has attracted increasing interests in the past decade. Evidences from various human behavior, ranging from E-mail communications [1] and [2], mobile communications [3] and [4], Web surfing [5], [6] and [7] to MicroBlog visiting [8] and [9] have shown that human dynamics are non-Poissonian, with bursts of active behavior separated by long periods of inactivity, which leads to a power law heavy tail of interval time between two consecutive behaviors [10]. Previous researches were conducted either on the collective level or the individual level, but the behavior dynamics during different active periods are still not well understood. In this paper, we focus on the human behavior during a special period called Flash Crowd (FC) period. FC is a large surge in traffic to a particular website causing a dramatic increase in server load [11]. Unlike network attacks, FC is an unintended phenomenon occurring as a consequence of collective reaction to a hot event. However, traffic fluctuation caused by FC is similar as that by distributed denial of service (DDoS) attack. Therefore, revealing human behavior dynamics in the FC period has important practical significance to network management. To uncover the underlying mechanism of human behavior during the FC period, we use data selected from World Cup 98 dataset which contains an FC. First, we provide approximate entropy (ApEn) to locate the FC stage from the request number sequence precisely. Then the multiscale entropy (MSE) method is used to analyze the difference of behaviors between the FC and Normal stages. Variation of entropy value between these stages may confirm the distinction of behavior dynamics. To reveal the origin of FC formation, we try to discuss the possible causes from the collective level and the individual level. After dividing the individuals into three groups according to their requests number of two stages, the distinction of behavior of the three groups in the FC stage is studied by MSE. We find that behavior in Group 2 contributes most to forming FC. Interval time distributions of request behavior from all individuals are further investigated. The majority of individual’s interval time distributions in Group 2 are more likely to follow Generalized Pareto (GP) compared to the Generalized Extreme Value (GEV) in Group 3. Finally, we study the changing of individual’s interval time distribution from the FC to the Normal stage in Group 2. The paper is organized as follows. In Section 2, we describe the dataset used in our empirical analysis. The methods we used are introduced in Section 3. The empirical results of behavior dynamics in the FC stage are presented in Section 4. In Section 5, we discuss the possible causes of FC formation, and the final section gives the conclusion.

In conclusion, we have studied the human behavior during a special period called FC in web surfing. The ApEn method was provided to separate the request number sequence into FC and Normal stages. We analyzed the difference of behaviors between these two stages by the MSE method. The lower entropy value in the FC stage may imply the more consistence of request behavior, which may be caused by the same purpose of users since the semi-final was about to begin. By dividing individuals into three groups, we discuss the possible causes of FC from the collective level. The results suggest that the origin of FC formation is not due to the increasing number of users but more likely to the change of behavior. Further investigation was conducted in the individual level. The majority of individuals’ interval time distributions have changed between two stages, which may further suggest that the change of dynamic mechanism of individual behavior is the cause to forming FC. However, it should be noted that our research was conducted on a single FC in one dataset. We need to do more investigations about different kinds of FC to make sure the key features we have found in this paper are general. And dynamic models of behavior will be constructed based on those common features in our future work.