Section: New Results

Characterizing Home Device Usage From Wireless Traffic Time Series

Participants: Katsiaryna Mirylenka (IBM), Vassilis Christophides, Themis Palpanas (University Rene Descartes), Ioannis Pefkianakis (HP Labs), Martin May (Technicolor)

We conducted a thorough analysis of traffic dynamics of heterogeneous wireless (WiFi) devices connected to 196 real RGWs, which are subscribers of a major European ISP. We focus on a time-oriented analysis of continuous traffic data to extract previously unknown patterns recurring of internet consumption that happen within, or across homes. We also assess the impact of different types of devices, such as laptops, desktops (classiffied as fixed devices), and tablets, smartphones (classiffied as portables), on these patterns. Unsupervised learning techniques are used for patterns discovery as the ground truth data regarding home activities are not available. Rather than partitioning homes or devices into distinct behavioral clusters, we are looking to extract informative motifs of bandwidth consumption within or across homes. The main contributions of this work are:

• We propose a novel analysis framework for wireless home traffic data, namely: (a) a correlation-based similarity measure, which exploits the evolution characteristics, rather than the absolute traffic values, and is invariant to scaling;(b) a notion of strong stationarity that in addition to the similarity of data distributions imposes a correlation similarity across non-overlapping time windows; and (d) a definition of dominant devices based on the correlation similarity, that enables an intuitive and statistically grounded interpretation of the results.

• We evaluate the effectiveness of the proposed framework using real data of wireless traffic observations and report the main findings: (a) there are many repetitive patterns within and across RGWs which describe the intrinsic user behavior of users and valuable to ISPs; (b) as networking time series are not stationary certain aggregation should be performed in order to find statistically significant patterns. The best time windows to aggregate home traffic data is found to be 8 hours for weekly patterns and 3 hours for daily patterns; (c) frequent weekly patterns correspond to heavy bandwidth usage both during weekdays and weekends, and frequent daily patterns correspond to (mostly) evening usage, (d) weekend usage tends to rely on portable devices, weekday usage relies more on fixed devices, while discontinuous usage within a day (mostly active in the evening or the morning) is still due to portable devices; and (e) almost every RGW involves a device that dominates its overall traffic, thus the behavior of this device should be mainly considered by ISPs while planning the updates.