The ubiquitous use of smart devices all over the world has contributed to an explosive increase in network traffic. Thus, methods for characterizing and predicting application-level traffic patterns from both individual and operator perspectives have received a lot of attention recently. Working along these lines, the authors of this paper make two important contributions: (1) they propose a systematic approach to group users into different profiles, along with identifying their traffic patterns at the application level; and (2) they propose an effective model for predicting application-level traffic usage of an individual.
With respect to (1), on a global level, the authors follow a well-known three-stage approach: feature extraction, feature reduction, and pattern identification. In the feature reduction step, they employ exploratory factor analysis borrowed from Mulaik [1]. With respect to (2), their hybrid model uses the wavelet transform [2] (borrowing ideas from [3]) and the auto regression and moving average (ARMA) approach [4]. This hybridization is particularly useful in the current context because the wavelet transform is useful for “break[ing] time sequences into a set of constitutive sequences,” and ARMA can “describe and predict time sequences with a properly fitted mathematical model.” Notably, their work is heavily influenced by the work of Furno et al. [5], which studies “network activity profiles and land usage from an operator perspective.”
The experiments were conducted on a dataset containing more than 350 thousand users’ complete application usage logs for six consecutive days in a week (only Wednesday was missing for reasons not mentioned). Thus, both weekday and weekend patterns can be studied. With the proposed approach, the authors identify six different user profiles. Among other things, the results include a key finding: a limited number of application-level traffic patterns exist among users, albeit with a skewed distribution therein. While this profiling is interesting and may be useful in future research, an appropriate validation step is missing here.
The authors’ wavelet-ARMA based model also shows promise when it comes to capturing the patterns of individual traffic dynamics. This is evident from the model’s median normalized mean square error (NMSE) of 0.1288. One limitation is the exclusion of weekend patterns, which may be of more interest to operators and content providers.