TomTom has the mission of creating a world free of congestion and better driving experience. In order to do that, we need to understand driving behavoiur from end users, at the same time that we optimize the operational costs of our services. However, due to the large scale of our probe data from vehicles providing insights and performing advanced analytics can can be quite challenging. During this discussion I will showcase two use cases where Databricks, Delta Lake and MLflow has enabled us to accelerate innovation. The first one is the IQMaps usecase. IQMaps is a system designed specifically for in-dash systems – taking the same up-to-date user experience you expect from navigation apps and bringing it to reliable, in-car navigation. IQ Maps learn the drivers’ driving patterns and updates the map regions that are most relevant to the user, using Wi-Fi or 4G. However, optimizing the data network consumption, which can have a high cost, while keeping the best driving experience, by having the map updated, requires complex simulations using millions of locations traces from vehicles. Apache Spark has been our key instrument to find the best balance to this trade off. The second use case is Destination Prediction. For many years, we have offered a personalized feature on our navigation products that predicts with high accuracy the driver’s next destination. Nonetheless, with the exponential increase and availability of data, and the access to more sophisticated Machine Learning models, we have revisited this feature to take it to the next level. Both us ecases take advantage of the latest frameworks and tools available on Databricks. With MLflow and Delta we have been able to find the best models that predict the destination for each individual driver, and to track each one of the KPIs.

1. WIFI SSID:Spark+AISummit | Password: UnifiedDataAnalytics

2. Sergio Ballesteros, TomTom
Kia Eisinga, TomTom
Driver Location Intelligence at
Scale using Apache Spark, Delta
Lake and MLflow on Databricks
#UnifiedDataAnalytics #SparkAISummit

3. Ourvision
A safe, connected, autonomous world that is free of
congestion and emissions.

4. 4
Bigdatadrivesour
business,but
dataprivacyalways
comesfirst

5. Data
• Anonymous location (GPS) traces
5

6. 742.000.000kmevery day
18.000 x
6

7. 7

8. Data
• Anonymous location (GPS) Traces
• Community inputs
• User events
• Journalistic data
• Car sensor data
8

9. Dataflow
9
~150 trillion
data points
~80 billion data
points per day

10. Dataflow
10

11. Dataflow
11

12. In dash systems are outperformed by smartphones
The embedded systemis expected to be up-to-date, with no user interaction. And the most visible component of it is a
map.
Usecase1:IQMapsanalytics
12

13. Driversdonotupdatetheirmaps
Today’s solutions provide manual updates,
oftenwith a necessity to drive to the dealer.
This is way too complex and inefficient.
13

14. 14

15. OEMsrequire dataefficient
solutions
While drivers expect up-to-date system, the carmakers
are usually concerned about the data cost required for
the map management.
15

16. 98% OF TRIPS ARE DRIVEN WITHIN150KM RADIUS99.8% OF TRIPS ARE DRIVEN WITHIN1000KM RADIUS
16

17. Whenradiusis0km
• User drives within 2 regions every week day
• Radius of 0 km.
• Download and install justhome regions
• Cellular data usage kept to a minimum
17

18. Whenradiusis150km
• User drives within 2 update regions every
week day
• Radius of 150 km.
• Home region: 6 update regions.
• Cellular data usage increased
18

19. IQMapsdemowithMLflow
19

20. 20

21. Realresultsusing0.5Mtrips
21
“This insight has led me to the conclusion
that a default radius of 150km is
unnecessary, and a small radius of ~10km
would already satisfy mostdrivers while
keeping cellular data usage low for OEMs.”
- Rolf Dorland, PM at TomTom

22. Goingonholidays
• User goes for his holiday (less frequent
updated region)
• Once user starts driving, updates for all
update regions the route goes through are
downloaded and installed.
22

23. 23
Destinationprediction

24. 24

25. Opportunity
25
Past: Rule-based solution
Delta Lake pipelines
Present: Machine Learning

26. Data
26
Original trace data from 1 source
227K device serials
Filtering out invalid trips
143K device serials
Users with at least 50 trips
3.6K device serials
Devices feasible for modelling
2.5K device serials

27. Features
For each trip, we have the following information:
• Where did the trip start?
• At what speed were you driving when the trip started?
• What was the time of day (morning/afternoon/evening) when the trip started?
• Was it rush hour when the trip started?
• What day of the week was it?
• Was it a weekend day?
• What was the season?
• Which driver profile do you belong to?
Historical information:
• Which destination did you go to your last trip? And the one before that? And the one before that?
• If it is a, let's say Monday, where did you go to the last Monday you made a trip? (do this for every weekday)
To predict: To which destination are you going?
What do we use in the end?
27

28. Labels
• We are given the latitude and longitude of a destination
of a trip.
• In order to find out which latitude and longitudes belong
to the same destination, we apply a clustering algorithm
called DBSCAN.
• DBSCAN clusters together destinations that are within
500 meters from each other. We should have at least 5
trips to a destination in order to call it a cluster.
How do we define where you are going?
28

29. 29

30. Train,validationandtestsplit
Trip ID Date Destination
Trip 1 January1 Cluster1
Trip 2 January22 Cluster2
Trip 3 February3 Cluster1
Trip 4 February15 Cluster2
Trip 5 March 2 Cluster1
Trip 6 March 14 Cluster1
Trip 7 March 27 Cluster2
Trip 8 April 4 Cluster1
Trip 9 April 16 Cluster2
Trip 10 May 8 Cluster1
Train
& validation
dataset
Test
dataset
TIME-SERIES CROSS-VALIDATION
Iterativeevaluation of the trips to
avoid overfitting
Trip ID Date Destination
Trip 1 January1 Cluster1
Trip 2 January22 Cluster2
Trip 3 February3 ?
Trip ID​ Date​ Destination
Trip 1​ January1​ Cluster1
Trip 2​ January22​ Cluster2
Trip 3​ February3​ Cluster1
Trip 4​ February15 ?
Data for 1 driver:
Trip ID​ Date​ Destination
Trip 1​ January1​ Cluster1
Trip 2​ January22​ Cluster2
Trip 3​ February3​ Cluster1
Trip 4​ February15 Cluster1
… … …
Trip 10 May 8 ?
30

31. Rapidexperimentation
31

32. 32

33. Majoritybaseline
Distribution of precision on the test set with a majority baseline classifier
33

34. Results
Distribution of precision on the test set with a tuned classifier
34

35. AcceleratingtheFutureofMobility
By embracing Apache Spark, Databricks and the Azure cloud
3535

36. DON’T FORGET TO RATE
AND REVIEW THE SESSIONS
SEARCH SPARK + AI SUMMIT