Swisscom is the leading mobile-service provider in Switzerland, with a market share high enough to enable us to model and understand the collective mobility in every area of the country. To accomplish that, we built an urban planning tool that helps cities better manage their infrastructure based on data-based insights, produced with Apache Spark, YARN, Kafka and a good dose of machine learning. In this talk, we will explain how building such a tool involves mining a massive amount of raw data (1.5E9 records/day) to extract fine-grained mobility features from raw network traces. These features are obtained using different machine learning algorithms. For example, we built an algorithm that segments a trajectory into mobile and static periods and trained classifiers that enable us to distinguish between different means of transport. As we sketch the different algorithmic components, we will present our approach to continuously run and test them, which involves complex pipelines managed with Oozie and fuelled with ground truth data. Finally, we will delve into the streaming part of our analytics and see how network events allow Swisscom to understand the characteristics of the flow of people on roads and paths of interest. This requires making a link between network coverage information and geographical positioning in the space of milliseconds and using Spark streaming with libraries that were originally designed for batch processing. We will conclude on the advantages and pitfalls of Spark involved in running this kind of pipeline on a multi-tenant cluster. Audiences should come back from this talk with an overall picture of the use of Apache Spark and related components of its ecosystem in the field of trajectory mining.

1. Mobility Insights at Swisscom :
Understanding Collective Mobility
in Switzerland
Spark Summit, October 2016
francois.garillot@swisscom.com @huitseeker
mohamed.kafsi@swisscom.com @mou7

2. Agenda
Intro
Smart-Data
Big Data Architecture
Trajectory Classification
Streaming
Data challenges

3. Introduction : Positioning

4. Positioning users in a modern
network
no triangulation at scale
positioning based on cell attachement history, prec ~200m
cell-to-cell handover, prec ~50m around limit
Timing Advance (roundtrip) : better results on good data sources

5. Trajectory data
mining
time series reconstruction
trajectory segmentation
map matching, clustering
mode of transport detection
...

6. How to create value with
positioning at Swisscom ?
with competitive analytics & data sources,
and by making sure it embodies the right values.

7. Smart Data

8. On (not) tracking (any users)
"Swisscom strictly complies with all applicable legislations, in
particular with the telecommunications law and the data
protection initiative."
Jürg Studerus, Swisscom Senior Manager, Corporate Responsibility

9. Smart Data : Big Data without Big Brother
Privacy preservation is an asset
It makes sense to care as much about your customer as they do about you.
We technically enforce this
answering only synoptic questions, no individual ones,
with data flow control : we neutralize quasi-identifiers at every stage

10. Swisscom mobile subscribers
source: xavierstuder.com, MD&A reports

11. Our choices
public good applications: making Switzerland run better,
understanding places, not individuals,
anonymized aggregations

12. A first product : City
"It's a dream for civil engineers" -- Alexandre Machu, Urban
systems engineer, Pully

13. Demo time

14. Usages
New roads to divert transit traffic out of downtown (informs a 50M$
project)
Parking lot expansion and transformation (informs a 10M$ project)
Electric car charging station deployment

15. Big Data architecture

16. In the backend

17. Spark configuration essentials for enterprise
jobs
spark.executor.memory="not the default 1g"
spark.kryo.registrator="something custom" // among others
spark.shuffle.service.enabled="true"
spark.dynamicAllocation.enabled="true"
spark.deploy.recoveryMode="ZOOKEEPER"
spark.deploy.recoveryDirectory="/path/to/state"
spark.deploy.zookeeper.url="quorumMachine1:2181, ..."
NOT the only valuable settings, see
https://techsuppdiva.github.io

18. Scala (1/2)
type ChronoHistory = List[UEupdate] @@ Chronological
type AnteChronoHistory = List[UEupdate] @@ AnteChronological
implicit class Chrono(l: List[UEupdate]) {
def asChrono: ChronoHistory = {
chronoCheck(l)
l.asInstanceOf[ChronoHistory]
}
def asAnteChrono: AnteChronoHistory = {
anteChronoCheck(l)
l.asInstanceOf[AnteChronoHistory]
}
}

19. Scala (2/2)
implicit def reverseChrono(l: ChronoHistory): AnteChronoHistory = l.reve
implicit def reverseAnteChrono(l: AnteChronoHistory): ChronoHistory = l.

20. >
Trajectory Classification

21. What is the proportion of trips
associated with trains?

22. Mode of Transport Detection
Input: Sequence of network events
Output: Mode of transport (train vs. other)
Network events associated with cells
Create fingerprints of cells
Intuition: cells with intermittent increases in the number of connections are
associated with collective mode of transports

23. Bursty Cell
Number of devices vs. minute of day

24. Burstiness
Random process with mean and variance , the relative variance isμ σ
2
.D =
σ
2
μ

25. Machine Learning with Spark
Periodic Spark job to compute cell features
Supervised training on labeled data (train vs. others)
Training and test with Spark ML

26. Spark (1/2)
val labeledPoints: RDD[LabeledPoint] = data.map {
case (transportMode, tripFeatures) =>
LabeledPoint(
labelOf(transportMode).toDouble,
featuresToFeatureVector(tripFeatures)
)
} // generate labeled data
labeledPoints.cache()
def trainNewModel = // Fix the used model
new LogisticRegressionWithLBFGS()
.setIntercept(true)
.setNumClasses(numberOfClasses)
.run(_: RDD[LabeledPoint])

27. Spark (2/2)
// train a model for performance evaluation
val model = trainNewModel(trainingData)
// Evaluate model on test instances and compute test error
val labelAndPreds = testData.map { point =>
val prediction = model.predict(point.features)
(point.label, prediction)
}
val testErr =
(labelAndPreds
.filter(r => r._1 != r._2)
.count().toDouble) / testData.count()
// train final model on the whole dataset
val finalModel = trainNewModel(labeledPoints)

28. Streaming Analytics

29. Road conditions on highways

30. Selecting users on a path of Interest

31. Graph matching
Locality-sensitive hashing :
A family H of hashing functions is
-sensitive if:
if then
if then
More:
LocalitySensitiveHashingBySpark, Uber, Spark Summit
2016
AGentleIntroductiontoLocality-SensitiveHashingwith
ApacheSpark, Scala By The Bay 2015
(r, cr, , )p1 p2
p– q ≤ r
P [h(q) = h(p)] ≥rH p1
p– q ≥ cr
P [h(q) = h(p)] ≤rH p2

32. Computing speeds: Solving graph
constraints
given a history of cells, where was the user, exactly ? (twice)
what's the path between 2 positions ?
linear query per user

33. Checkpointing: Set the checkpoint
interval
are you checkpointing too often ?
every batches, you'll need batches to recover from checkpointing time
loss
make sure
k p
k ≥ p

34. Data Challenges

35. Crucial elements
Quality, reliability of data sources
Automated ground truth checking
sensors
TEMS fleet
What's the ground truth for mode of transport, domicile, etc ?
Colleagues and friends volunteers

36. Questions ?