Observing Consistency in Online Communication Patterns for User Re-Identification.

Ikuesan Richard Adeyemi,Shukor Abd Razak,Mazleena Salleh,Hein S Venter,Sergio Gómez

doi:10.1371/journal.pone.0166930

Ikuesan Richard Adeyemi, Shukor Abd Razak + Show 3 more

Open Access

https://doi.org/10.1371/journal.pone.0166930

Copy DOI

Abstract

Comprehension of the statistical and structural mechanisms governing human dynamics in online interaction plays a pivotal role in online user identification, online profile development, and recommender systems. However, building a characteristic model of human dynamics on the Internet involves a complete analysis of the variations in human activity patterns, which is a complex process. This complexity is inherent in human dynamics and has not been extensively studied to reveal the structural composition of human behavior. A typical method of anatomizing such a complex system is viewing all independent interconnectivity that constitutes the complexity. An examination of the various dimensions of human communication pattern in online interactions is presented in this paper. The study employed reliable server-side web data from 31 known users to explore characteristics of human-driven communications. Various machine-learning techniques were explored. The results revealed that each individual exhibited a relatively consistent, unique behavioral signature and that the logistic regression model and model tree can be used to accurately distinguish online users. These results are applicable to one-to-one online user identification processes, insider misuse investigation processes, and online profiling in various areas.

Highlights

Based on a series of empirical validations and several theoretical assumptions, the Internet is asserted to display fractal behavior with structural similarities irrespective of timescale
With reference to the empirical observation in [10], where the decay of visitation patterns correlates with the visitation pattern, this study revealed that the length of a session is significantly related to the visitation patterns and aggregated visitation patterns
The study attempted to address two pertinent research questions on online attribution: Is user behavior consistent over the Internet? If it is, to what extent can the distinction be applied to distinguish online users? The dynamic characteristics of individual users were initially observed and subsequently applied to the exploration of the probability of the existence of consistent online browsing patterns using the fundamental unit of client-server communication processes

Summary

Introduction

Based on a series of empirical validations and several theoretical assumptions, the Internet is asserted to display fractal behavior with structural similarities irrespective of timescale. This observation is similar to that found in [1,4], where online behavior is asserted to obey the power law. This is achieved through information gain principle (which measures the expected reduction in entropy), tree pruning based on reduced-error pruning with the back fitting method, and integration of C4.5 mechanism for missing value by splitting each corresponding instances into fractional instances Both logistic regression model and LMT demonstrated higher classification accuracy on the test sets than REPTree. A uniform class prior probability is not a precursor to online user reidentification study since the presence of the behavioral pattern is not dependent on an equal number of sessions for each observed User

Research Limitations and Future Opportunities

Findings

Conclusions