1 of 41

The WWW �(and an H) of Mobile Application Usage in the City

E. Graells-Garrido, D. Caro, O. Miranda, R. Schifanella, O. Peredo�Data Science Institute, U. del Desarrollo / U. de Chile / U. of Turin / Telefónica R&D

Stanley Kubrick. 1946.

2 of 41

The urban context (and our ability to carry things) certainly influences what we do.

Walter Evans, 1938

Stanley Kubrick, 1946.

3 of 41

Motivation

Just by carrying a smartphone we can do almost everything!*�

*your mileage may vary

https://en.wikipedia.org/wiki/Smartphone#/media/File:Fotos_produzidas_pelo_Senado_(30554309793).jpg

4 of 41

Motivation

The Web has become mobile!

Research has focused on:

Informational and leisure tasks in mobile context
Predicting next app to be opened
Understanding location context

5 of 41

Problem

��Little is known regarding how the urban fabric and the activities that take place in it affect the usage of mobile applications.

https://www.flickr.com/photos/greenlaneproject/22729109455

6 of 41

Proposal

A descriptive study of:

What are the most popular applications used in the city?

Where are they spatially clustered?

When an application usage is more frequent?

and

How does the urban context and the mobility patterns relate to application usage?

https://commons.wikimedia.org/wiki/File:2013_people_wating_on_subway_platform_12677922605.jpg

7 of 41

Context: �Santiago, Chile

@rocco.jpg

8 of 41

Santiago Metro. Area: ~8 million inhabitants (largest metropolitan area, with almost half of the country’s population)

Grand Lyon: ~1.35 million inhabitants. (second largest metropolitan area, third largest city in France)

9 of 41

Economic Segregation (Left). Population Distribution (Center) and Work Areas (Right). Each dot represents a trip origin/destination.

10 of 41

In 2016, there were 132 mobile subscriptions per 100 inhabitants in Chile.

11 of 41

Datasets

12 of 41

Deep Packet Inspection (DPI)

Used for Quality of Service (QoS)

Mobile Operator: Telefonica Movistar Chile (33% market share)

Period: from 27th July to 10th August, 2016 (15 days)

# of Requests to the most popular 5,000 IPs in time windows of 10 minutes, per tower

13 of 41

Mobile Phone Towers

Mobile phone towers have:

Unique identificator.
Geographical position.
Indoor label (True/False).
Name (used to identify towers within underground Metro stations, and towers within tunnels and malls).

14 of 41

OpenStreetMap

We use an OSM dump to find urban infrastructure relevant for the study, such as:

Highways
Primary and Secondary Streets
Pedestrian Streets
Bus Corridors
Green Areas
Surface Metro routes

15 of 41

Results: The What

16 of 41

Apps and Websites

We mapped the 5,000 IPs to specific websites and applications by whois, manual lookup into domain databases, and security-related databases.

In total, we identified 1133 IPs.

The others were:

DNS or ISP related
Ambiguous (for instance, map providers for several apps)
Generic (Amazon Web Services)
Ads and Analytics

17 of 41

Results: The Where

18 of 41

Methods

To understand where applications are used, we measured Getis-Ord Spatial Autocorrelation metrics.

Global Autocorrelation: is the app used in clusters that are more concentrated than a random model?�
Local Autocorrelation: hotspots of the city (groups of towers) where two (or more) nearby towers tend to use the same app with a greater volume than other groups of towers in the city.

19 of 41

Global

Most of the applications exhibit some kind of clustering or dispersion at some parts of the day (The grey band represents a random model of usage).

Many apps show global autocorrelation in the afternoon, due to their concentration in work/study areas.

Facebook is so popular that is more dispersed than in a random model at any time.

More clustered than in a random model

More dispersed than in a random model

20 of 41

Local autocorrelation results.

Dots are the hotspots of each app.

21 of 41

Results: �The When �(and How)

22 of 41

The When �(and How)

mean # requests per tower.

To evaluate the how, we performed several Negative Binomial Regression with several urban factors, including:

Indoor and Metro Stations
Bus Rapid Transit, Highways, and Pedestrian Streets
Street Importance (Primary, Secondary)
Income
Green Areas

(see the paper for model details)

23 of 41

Regression Results

(See the paper for details 😱)

Let’s look at individual apps!

24 of 41

Music

Music is mostly used at commuting times (see Intercept).

Its association to metro stations is high (with an effect size greater than 400% increase with respect to the base), particularly in the mornings.

Confidence Interval

25 of 41

Games

Games are used mostly in transportation infrastructure: metro stations, bus corridors, and within metro stations.

Their usage increases during the day.

26 of 41

Instagram

As with other apps, Instagram is heavily used in transportation infrastructure, particularly public transportation.

The patterns behind metro usage may indicate that Instagram is used by people that have both, half-time and full time work/study journeys, due to the appearance of a third peak at lunch time.

We expected influence of green areas, but found opposite evidence! 💔

27 of 41

Twitter

As with other apps, Twitter tends to be used in commuting times.

In contrast with music, its association with metro is almost symmetric with respect to the rush hours.

It also shows a noticeable effect in primary streets and bus corridors.

28 of 41

Dating

These applications show a small but steady increase of usage during the day.

They don’t exhibit strong associations except with:

Metro at lunch times (perhaps by a younger audience)
Pedestrian streets at dinner / night time (perhaps by a more adult audience, or the availability of meeting points). Since these streets are close to primary streets, the effect permeates.

29 of 41

Discussion �and Conclusions

30 of 41

A Descriptive Study

Commuting is relevant to what users do on their phones!

Commuting times are on the rise! So this will be more important in the future.
Commuting methods are also likely to change (automated cars?)

Many known patterns, now quantified.

Information consumption: people read (and maybe post) at commuting times and breaks.
Facebook so popular that is more dispersed than random models.

Applications that are perceived as biased (e.g., Twitter), are used through the city.

Perhaps the bias is more on user-generated content, and less on usage.
Residential Income was not a relevant factor (and when it was, effect size was very small).

We expected some results but found that the opposite happens (e.g., Instagram).

31 of 41

Implications and Future Work

Inspecting packet data makes it possible to infer many things about the population.
What are the relationships of people and places (see our other works) with informational behavior?

Are apps a replacement of POIs or an alternative to the lack of opportunities? (i.e., does people watch video because there is a lack of cinemas?)

Country perspective -- Chile has an interesting mixture of geographical and weather patterns.

The most arid desert in the north, forest in the south, and ice and wind in the extreme south.

And other questions you may have... let’s collaborate :)

32 of 41

Thanks!

Any questions?

Write us! �egraells@udd.cl / @carnby