Source-linked AI summary
Human dynamics revealed through Web analytics
Bruno Goncalves, Jose J. Ramasco
TL;DR
The paper asks whether Web use exhibits systematic human-dynamics patterns and analyzes anonymized, consistently trackable Emory University server logs. It finds structured population activity and argues that priority queuing, linear preferential linking, and Web-page aging are essential factors in modeling Web surfing.
Problem
The paper investigates how people interact with the Web and whether visiting frequency follows shared laws or remains intrinsically individual.
Method
The authors analyze anonymized Emory University Web-server logs, tracking users consistently while examining population activity, individual inter-click times, preferential linking, and site dynamics.
Results
The study identifies structured activity patterns and finds that priority queuing, linear preferential linking, and Web-page aging are essential factors for characterizing Web surfing.
Takeaways & Limitations
Web-server logs can reveal human activity rhythms and support models combining priority queuing, preferential linking, and Web-page aging.
Abstract
from arXiv · showhide
When the World Wide Web was first conceived as a way to facilitate the sharing of scientific information at the CERN (European Center for Nuclear Research) few could have imagined the role it would come to play in the following decades. Since then, the increasing ubiquity of Internet access and the frequency with which people interact with it raise the possibility of using the Web to better observe, understand, and monitor several aspects of human social behavior. Web sites with large numbers of frequently returning users are ideal for this task. If these sites belong to companies or universities, their usage patterns can furnish information about the working habits of entire populations. In this work, we analyze the properly anonymized logs detailing the access history to Emory University's Web site. Emory is a medium size university located in Atlanta, Georgia. We find interesting structure in the activity patterns of the domain and study in a systematic way the main forces behind the dynamics of the traffic. In particular, we show that both linear preferential linking and priority based queuing are essential ingredients to understand the way users navigate the Web.
I. INTRODUCTION
The paper asks whether Web-visiting behavior follows shared statistical laws or remains intrinsically individual. It addresses this by analyzing anonymized, individually trackable requests to Emory University's Web domain.
- The study examines whether laws govern how and how often people visit a given Web site, rather than each person's behavior being entirely unique.
- Emory Web-server logs provide an open-environment dataset covering internal and external requests to second-level domain pages from April 1, 2005, to January 17, 2006.
- Privacy-preserving anonymization replaces real identities with consistent IDs, allowing each user's behavior to be followed without revealing the person's identity.
II. ACTIVITY PATTERNS OF THE POPULATION
Emory Web traffic varies systematically across weeks, holidays, and hours, reflecting the university's academic and work rhythms. The data show weekly and daily activity patterns, including Thanksgiving effects and reduced Sunday activity.
- A weekly oscillation and distinct late-August and late-December trends align Web traffic with the beginning and end of the academic year.
- Sundays are the least active day, while activity generally declines after Wednesday; traffic is treated as an indirect measure of university productivity because access is mostly work related.
- During Thanksgiving week, activity falls across the break, Mondays before and after are less productive, and Tuesdays surrounding the break are busier than usual.
- Weekday activity is highest from 7AM to 6PM, dips from 11AM to 2PM during lunch, peaks again afterward, and levels off after 6PM.
- Saturdays resemble other days in hourly profile, whereas Sundays show a distinct profile consistent with reported human circadian patterns.
III. INDIVIDUAL ACTIVITIES
Individual Web activity varies widely across users and includes distinct machine-like and human-like patterns. Inter-click times show broad distributions that can be modeled with priority-based task-queue dynamics, including a multi-task extension for domain-wide clicks.
- User heterogeneity: Wide user-activity distributions make averages poor descriptors because most users generate few requests while some accumulate many.The paper notes that some single-user activity distributions are well modeled by power laws with exponents below 2.
- User heterogeneity: Activity profiles distinguish an apparent malicious attack, an automatic software updater, and a human entering data on an administrative site.The attack concentrates many regular requests in a short interval, the updater visits regularly, and the human profile differs from both machines.
- Inter-click dynamics: Inter-click times are broad for both repeated visits to one URL and visits to any Emory page, with the latter modeled by a power law of exponent −1.25.Figure 6 compares same-user same-URL intervals with same-user domain-wide intervals.
- Inter-click dynamics: Repeated visits to the same URL are associated with a power-law exponent of −1, which the authors relate to a priority-based task model.In that model, agents select the highest-priority task with probability p and a random task otherwise.
- Inter-click dynamics: Selecting groups of ν tasks per step explains the domain-wide exponent −1.25 when ν > 2, although the proposed models are acknowledged as oversimplifications.The paper reports a systematic numerical study and illustrates ν = 3 in the inset of Figure 6b.
IV. ATTRACTIVENESS AND PREFERENTIAL LINKING
The study finds linear preferential linking across user–Web-page interactions, while URL lifetimes tied to the academic year complicate longer-term distributions.
- IV. ATTRACTIVENESS AND PREFERENTIAL LINKING: A systematic study of preferential linking is used to test whether new connections and clicks favor already popular Web elements.The analysis considers Web pages and IPs as system elements.
- IV. ATTRACTIVENESS AND PREFERENTIAL LINKING: Linear preferential functions describe weekly changes in user exploration, URL attraction of new users, and revisits to Web pages.The variables are kIP, kURL, and w, respectively.
- IV. ATTRACTIVENESS AND PREFERENTIAL LINKING: For C(kIP), short- and long-window distributions overlap and fit a power law with exponent γ ≈2.2, unlike C(kURL) and w.The comparison uses one-week accumulation with the full 292-day database.
- IV. ATTRACTIVENESS AND PREFERENTIAL LINKING: URL creation and extinction peaks coincide with semester beginnings and endings, indicating that page relevance is linked to the academic-year cycle.Many pages therefore have relatively short lifetimes and may no longer be maintained after the semester.
V. DISCUSSION AND CONCLUSIONS
Using 292 days of anonymized Emory Web-server logs, the paper connects university traffic to human activity patterns and models user navigation across scales. It concludes that priority queuing, preferential linking, and Web-page aging are jointly important for characterizing Web surfing.
- V. DISCUSSION AND CONCLUSIONS: The study identifies human circadian rhythms and activity patterns consistent with prior mobile-phone and social-email observations in university Web traffic.The logs cover 292 days and include students, professors, and staff.
- V. DISCUSSION AND CONCLUSIONS: Priority-queue models explain inter-click periods for user–site pairs, while selecting multiple tasks at once accounts for whole-domain request intervals.The modified model executes more than one task in a short interval.
- V. DISCUSSION AND CONCLUSIONS: Linear preferential attractiveness appears in user exploration, sites’ acquisition of visitors, and new clicks on user–Web-page connections over short periods.Longer periods require accounting for Web-page lifetimes.
- V. DISCUSSION AND CONCLUSIONS: Priority queuing, preferential linking, and Web-page aging are identified as essential factors for models of Web surfing.This is the paper’s concluding model-level synthesis.