Authors: Andrea Hess, Ian Marsh and Daniel Gillblad (SICS Swedish ICT)
Venue: IEEE International Conference on Communications (ICC), London, June 2015, pp. 5916–5921
DOI: 10.1109/ICC.2015.7249265
Year: 2015
Topic: Big data and analytics
Summary
Mobile operators increasingly use data analytics to plan their networks. Earlier studies mostly looked at the network as a whole: where and when the peak loads occur. Studies that classified users typically put each subscriber into one group, such as "heavy streamer" or "commuter", and left them there.
People do not stay in one group. A student's behaviour changes between term and holiday; a commuter's changes when they switch jobs. This paper proposes a profiling model that follows individual users over time, along two dimensions:
- Communication: the network load a user generates (sessions, volumes, timing).
- Mobility: how a user moves through the network (how many cells they visit, how far, how regularly).
The model was tested with two unsupervised clustering algorithms, X-means and expectation maximisation (EM), on a 3G trace from a European mobile operator. The main question is temporal consistency: how stable is a user's profile from one period to the next, and which users move between clusters?
Approach
The model has three steps. (The examples below are typical of this kind of study; the paper has the exact feature set.)
- Feature extraction. For each user and time window, compute load features from the trace (for example traffic volume, number and length of sessions, time of day) and mobility features (for example the number of distinct cells visited and how regularly the user returns to them).
- Clustering. Group users in each window. X-means is k-means that picks the number of clusters itself using the Bayesian information criterion. EM fits a Gaussian mixture, so each user gets a probability of belonging to each cluster rather than a hard label.
- Tracking. Compare a user's cluster across consecutive windows, and measure how much the clusters themselves change.
Clustering the two dimensions separately and together shows whether heavy users are also highly mobile, and whether load or mobility is the more stable trait.
Why it matters
- Network planning. If a group of users is consistently heavy and mobile, capacity has to follow them along their routes, not only sit where peak loads have been measured.
- Resource allocation and caching. Predictable users can have content prefetched or radio resources reserved ahead of time.
- Anomaly detection. A user who suddenly leaves their usual cluster may have a faulty device, a stolen SIM, or a changed life.
- Products. Tariffs and services can follow how people actually use the network.
Discussion
Clustering mobile users raises two methodological questions that this paper puts at the centre:
- How many groups are there? Fixing k in advance shapes the answer. X-means and EM both choose the model size from the data, but with different assumptions: hard boundaries against overlapping, probabilistic ones.
- Are the groups real? A clustering that changes completely from week to week says more about noise than about people. Measuring temporal consistency is a test of whether the profiles are worth acting on.
The work belonged to a wider SICS effort on large-scale mobile data. Related work on this site includes Delay characterisation through in-protocol measurements (mobile delays seen in TCP traffic) and Dimensionality reduction in (large) measurement datasets, which looks at the same kind of clustering problem from the visualisation side.
Future work
- From 3G to 5G. 5G standalone networks expose much richer per-user data through the Network Data Analytics Function (NWDAF, 3GPP TS 23.288), which is designed to produce exactly this kind of user and mobility analytics for the network itself.
- Online profiles. Update clusters continuously in a stream rather than in batch windows, and raise an alert when a user changes profile.
- Sequence models. Predict a user's next cluster, or next cell, with recurrent or transformer models trained on the profile history.
- Privacy-preserving profiling. Federated learning or differential privacy, so that profiles can be built without collecting raw location traces centrally.
The EU perspective
Mobile traces are among the most sensitive data an operator holds, and EU law treats them that way:
- GDPR: subscriber-level location and usage data are personal data. Profiling is explicitly covered (Article 4(4)), and automated decisions with significant effects need safeguards (Article 22). Mobility traces are notoriously hard to anonymise: four location points are enough to identify most people in a large dataset (de Montjoye et al., 2013).
- ePrivacy Directive 2002/58/EC, Articles 6 and 9: traffic and location data may be processed for network management and billing, but other uses need user consent or anonymisation. The proposed ePrivacy Regulation was withdrawn by the Commission in 2025, so the Directive still applies.
- Official statistics: Eurostat and national statistics offices have explored mobile network operator data for population and tourism statistics. The Commission's 2023 proposal to revise the European statistics regulation would give statistical authorities access to privately held data such as mobile data.
- Pandemic use: in 2020 the Commission and European operators shared aggregated, anonymised mobility data to model the spread of COVID-19, under Recommendation (EU) 2020/518. This was a large-scale test of the kind of mobility profiling described here, under strict aggregation.
- AI Act (EU) 2024/1689: profiling users for network planning is not high-risk, but profiling used for credit scoring or access to essential services would be.
References
- A. Hess, I. Marsh and D. Gillblad, "Exploring communication and mobility behavior of 3G network users and its temporal consistency," IEEE ICC, 2015, pp. 5916–5921.
- D. Pelleg and A. Moore, "X-means: extending k-means with efficient estimation of the number of clusters," ICML, 2000.
- M. C. González, C. A. Hidalgo and A.-L. Barabási, "Understanding individual human mobility patterns," Nature 453, 2008.
- Y.-A. de Montjoye, C. A. Hidalgo, M. Verleysen and V. D. Blondel, "Unique in the crowd: the privacy bounds of human mobility," Scientific Reports 3, 2013.
- 3GPP TS 23.288, Architecture enhancements for 5G System to support network data analytics services.
- Commission Recommendation (EU) 2020/518 on a common Union toolbox for the use of technology and data to combat the COVID-19 crisis.
Citation
Andrea Hess, Ian Marsh and Daniel Gillblad, Exploring communication and mobility behaviour of 3G network users and its temporal consistency, IEEE ICC, 2015, Paper.
BibTeX
@inproceedings{hess2015exploring,
title = {Exploring communication and mobility behavior of {3G} network users and its temporal consistency},
author = {Hess, Andrea and Marsh, Ian and Gillblad, Daniel},
booktitle = {IEEE International Conference on Communications (ICC)},
pages = {5916--5921},
year = {2015},
doi = {10.1109/ICC.2015.7249265}
}
More in Big data and analytics: ← 12 years of traffic flow data | Data Readiness for Big Automotive Data Analytics (BADA) →