The surge in mobile data traffic -estimated globally at 3.7 exabytes in 2015, with a 74% increase over 2014 and an overall 4,000-fold growth over the past ten years- has fostered the interest of the computer network community towards better understanding the dynamics of the mobile demand. Indeed, a proper characterization of how mobile services are consumed by subscribers can enable an informed, more efficient tailoring of network resource planning and management to the end users' needs. Knowledge mining of real-world datasets has revealed important features of mobile traffic. Examples include a strong temporal periodicity and geographic locality that enable effective prediction of the demand; the appearance of significant fluctuations induced by social events with the consequent need for dedicated resource management policies; or, a neat heterogeneity of the capacity consumed by subscribers that is however captured by a limited number of typical profiles, which enables the informed tuning of traffic plans.
Our work focuses on the problem of classification, i.e., finding hidden regular structures in the network-wide aggregate traffic generated by mobile users. The problem can be cast in the temporal or spatial dimensions, both of which are relevant to network operation management.
In the temporal dimension, the problem is that of network activity profiling, i.e., classifying together time periods that show a similar, stable spatial distribution of the mobile traffic demand. Network activity profiles find applications in cognitive networking, as they can drive the establishment and relocation of resources in concert with the temporal variations in the mobile demand.
In the spatial dimension, the problem is that of land use detection, i.e., the decomposition of a geographical area into zones where the mobile traffic dynamics are homogeneous over time. These zones typically correspond to land uses, i.e., the combination of urban infrastructures and predominant undertakings of people at a location. This knowledge can support the dynamic allocation of capacity at individual base stations, and help mitigating high fluctuations of resource needs in small-sized network areas. In addition, land use detection has applications in geoinformatics, as an effective way to automatically label the urban tissue, at lower cost and with higher accuracy than traditional survey methods.
We present an original methodology for the joint spatiotemporal classification of the aggregate demand that a mobile net- work has to serve. Our solution stems from Exploratory Factor Analysis (EFA), a well-established instrument in psychology research. EFA aims at identifying, in a fully automated way, latent factors that cause the dynamics observed in the data. When tailored to the specific use case of mobile traffic classification, EFA offers the possibility of exploring the space and time dimensions of the data at once. Extensive tests with real-world data provided by major mobile network operators show that EFA allows immediate extrapolation of the structures hidden in the secondary dimension of each problem. In other words, it provides, at no additional cost, knowledge of the spatial patterns that characterize each network activity profile, and of the precise temporal dynamics that distinguish each land use. As an example, the figure shows the geographical areas in Milan that are the most relevant to the temporal profile determined by student nightlife activity.
More details in: A. Furno, M. Fiore, R. Stanica, Joint Spatial and Temporal Classification of Mobile Traffic Demands, IEEE INFOCOM 2017, Atlanta, GA, USA, April 2017.
Focus