ZipNet-GAN: Inferring Fine-grained Mobile Traic Paerns via a Generative Adversarial Neural Network Chaoyun Zhang Xi Ouyang Paul Patras School of Informatics School of Electronic Information and School of Informatics University of Edinburgh, UK Communications, University of Edinburgh, UK [email protected] Huazhong University of Science and [email protected] Technology, China [email protected] ABSTRACT 1 INTRODUCTION Large-scale mobile trac analytics is becoming essential to digital e dramatic uptake of mobile devices and online services over infrastructure provisioning, public transportation, events planning, recent years has triggered a surge in mobile data trac. Industry and other domains. Monitoring city-wide mobile trac is however forecasts monthly global trac consumption will surpass 49 ex- a complex and costly process that relies on dedicated probes. Some abytes (1018) by 2021, which is a seven-fold increase of the current of these probes have limited precision or coverage, others gather utilisation [1]. In this context, in-depth understanding of mobile tens of gigabytes of logs daily, which independently oer limited in- trac paerns, especially in large-scale urban cellular deployments, sights. Extracting ne-grained paerns involves expensive spatial becomes critical. In particular, such knowledge is essential for dy- aggregation of measurements, storage, and post-processing. In this namic resource provisioning to meet end-user application require- paper, we propose a mobile trac super-resolution technique that ments, Internet services that rely on user location information, and overcomes these problems by inferring narrowly localised trac the optimisation of transportation systems [2–4]. consumption from coarse measurements. We draw inspiration from Mobile trac measurement collection and monitoring currently image processing and design a deep-learning architecture tailored rely on dedicated probes deployed at dierent locations within the to mobile networking, which combines Zipper Network (ZipNet) network infrastructure [5]. Unfortunately, this equipment (i) either and Generative Adversarial neural Network (GAN) models. is en- acquires only coarse information about users’ position, the start ables to uniquely capture spatio-temporal relations between trac and end of data sessions, and volume consumed (which is required volume snapshots routinely monitored over broad coverage areas for billing), while roughly approximating location throughout ses- (‘low-resolution’) and the corresponding consumption at 0.05 km2 sions – Packet Data Network Gateway (PGW) probes; (ii) or has level (‘high-resolution’) usually obtained aer intensive computa- small geographical coverage, e.g. Radio Network Controller (RNC) tion. Experiments we conduct with a real-world data set demon- or eNodeB probes that need to be densely deployed, require to strate that the proposed ZipNet(-GAN) infers trac consumption store tens of gigabytes of data per day at each RNC [6], but cannot with remarkable accuracy and up to 100× higher granularity as independently quantify data consumption. In addition, context compared to standard probing, while outperforming existing data information recorded is oen stale, which renders timely infer- interpolation techniques. To our knowledge, this is the rst time ences dicult. Mobile trac prediction within narrowly localised super-resolution concepts are applied to large-scale mobile trac regions is vital for precision trac-engineering and can further analysis and our solution is the rst to infer ne-grained urban benet the provisioning of emergency services. Yet this remains trac paerns from coarse aggregates. costly, as it involves non-negligible overheads associated with the transfer of reports [5], substantial storage capabilities, and inten- CCS CONCEPTS sive o-line post-processing. In particular, combining dierent •Networks →Network measurement; information from a large number of such specialised equipment is required, e.g. user position obtained via triangulation [7], data KEYWORDS session timings, trac consumed per location, etc. Overall this is a challenging endeavour that cannot be surmounted in real-time by mobile trac; super-resolution; deep learning; directly employing measurements from independent RNC probes. generative adversarial networks In order to simplify the analysis process, mobile operators make simple assumptions about the distribution of data trac consump- tion across cells. For instance, it is frequently assumed users and Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed trac are uniformly distributed, irrespective of the geographical for prot or commercial advantage and that copies bear this notice and the full citation layout of coverage areas [8]. Unfortunately, such approximations on the rst page. Copyrights for components of this work owned by others than the are usually highly inaccurate, as trac volumes exhibit consid- author(s) must be honored. Abstracting with credit is permied. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specic permission erable disparities between proximate locations [3]. Operating on and/or a fee. Request permissions from [email protected]. such simplied premises can lead to decient network resources CoNEXT ’17, Incheon, Republic of Korea © 2017 Copyright held by the owner/author(s). Publication rights licensed to ACM. 978-1-4503-5422-6/17/12...$15.00 DOI: 10.1145/3143361.3143393 CoNEXT ’17, December 12–15, 2017, Incheon, Republic of Korea C. Zhang et al. inspires us to employ image processing techniques to learn end- Low-resolution image High-resolution image to-end relations between low- and high-resolution mobile trac snapshots. We illustrate the similarity between the two SR problems in Fig. 1. Recent achievements in GPU based computing [12] have Image SR led to important results in image classication [13], while dierent neural network architectures have been successfully employed to learn complex relationships between low- and high-resolution images [14, 15]. us we recognise the potential of exploiting deep learning models to achieve reliable mobile trac super-resolution Coarse-grained Measurements Fine-grained Measurements (MTSR) and make the following key contributions: (1) We propose a novel Generative Adversarial neural Network (GAN) architecture tailored to MTSR, to infer ne-grained mo- bile trac paerns, from aggregate measurements collected MTSR by network probes. Specically, in our design high-resolution trac maps are obtained through a generative model that out- puts approximations of the real trac distribution. is is trained with a discriminative model that estimates the prob- ability a sample snapshot comes from a ne-grained ground truth measurements set, rather than being produced by the generator. (2) We construct the generator component of the GAN using an Figure 1: Illustration of the image super-resolution (SR) original deep zipper network (ZipNet) architecture. is up- problem (above) and the underlying principle of the pro- grades a sophisticated ResNet model [16] with a set of addi- posed mobile trac super-resolution (MTSR) technique (be- tional ‘skip-connections’, without introducing extra parame- low). Figure best viewed in color. ters, while allowing gradients to backpropagate faster through the model in the training phrase. e ZipNet further introduces 3D upscaling blocks to jointly extract spatial and temporal fea- tures that exist in mobile trac paerns. Our design acceler- allocations and implicitly to modest end-user quality of experi- ates training convergence and we demonstrate it outperforms ence. Alternative coarse-grained measurement crowdsourcing [9] the baseline Super-Resolution Convolutional Neural Network remains impractical for trac inference purposes. (SRCNN) [14] in terms of prediction accuracy. In this paper we propose an original mobile trac ‘super-reso- (3) To stabilise the adversarial model training process, and prevent lution’ (MTSR) technique that drastically reduces the complexity model collapse or non-convergence problems (common in GAN of measurements collection and analysis, characteristic to purely design), we introduce an empirical loss function. e intuition RAN based tools, while gaining accurate insights into ne-grained behind our choice is that the generator can be optimised and mobile trac paerns at city scale. Our objective is to precisely infer complexity reduced, if using a loss function for model training narrowly localised trac consumption from aggregate coarse data that is insensitive to model structure and hyper-parameters recorded by a limited number of probes (thus reducing deployment conguration. costs) that have arbitrary granularity. Achieving this goal is chal- (4) We propose a data processing and augmentation procedure to lenging, as small numbers of probes summarise trac consumption handle the insuciency of training data and ensure the neural over wide areas and thus only provide ‘low-resolution’ snapshots network model does not turn signicantly over-ed. Our that cannot capture distributions correlated with human mobility. approach crops the original city-wide mobile data ‘snapshots’ Further, measurement instrumentation is non-uniformly deployed, to smaller size windows and repeats this process with dierent as coverage area sizes depend on the population density [10]. Such osets to generate extra data points from