Inferring land use from mobile phone activity

Size: px

Start display at page:

Download "Inferring land use from mobile phone activity"

Peregrine Banks
6 years ago
Views:

1 Inferring land use from mobile phone activity Jameson L. Toole (MIT) Michael Ulm (AIT) Dietmar Bauer (AIT) Marta C. Gonzalez (MIT) UrbComp 2012 Beijing, China 1

2 The Big Questions Can land use be predicted from mobile phone activity? Can mobile phone data help measure temporal patterns of land use? Are these patterns robust across cities? 2

3 Sensing Urban Systems Data Scarcity Methods Outputs Abundance Traditional Data: Zoning Maps Novel Data: Mobile phone activity Statistical Physics Machine learning Geographic Boundaries (Ratti et al 2010) Intra-city Mobility (Cho et al 2011) Inter-city Mobility (Gonzalez et al 2008) (Wang et al 2009) (Simini 2012) Unsupervised Classification (Reades 2009) (Calabrese 2010) (Soto 2011)(Zheng 2012) 3

4 Land use and Travel Behavior travel demand Land-use supply Temporal Component opening hours of shops and services available time Individual Component income, gender, education level Travel supply Demand available time for activities vehicle ownership, etc. Demand where a person wants to go when a person can travel a persons ability to travel a persons desire to travel Static GIS data + = New micro level data from mobile phones Dynamic Land Use adapted from: K.T. Geurs, B. van Wee / Journal of Transport Geography 12 (2004)

5 Data Zoning Data GIS shapefiles at the parcel level Provided by MASSDOT Aggregated to 5 zoning classification: Residential, Commercial, Industrial, Parks, Other Mobile Phone ~600,000 Users in the Boston Metropolitan Area (Pop: ~3 million, Mkt Share: 30%-50%) 1 month of Call Detail Records (CDR) for voice calls and text messages. Each event geo-tagged with (Lat, Lon) triangulated from nearby towers Grid Partition the region into 200m X 200m cells Each cell is labeled with the most prevalent zoning type within Grid provides privacy All mobile phone events are aggregated hourly into an average week 5

6 6

7 ZONING DATA 40km Parks Other CBD 7

8 Zoned use & Dynamic Population Results: Absolute Activity Downtown Boston is mostly zoned as OTHER thus more activity is related to more people. Abs. Act Parks Other 8

9 Zoned use & Dynamic Population Results: Normalized Activity a norm ij (t) = aabs ij (t) µ a abs ij σ a abs ij 1 0 Parks Other Similarities are related to how individuals use mobile phones, not how they move. 9

10 Zoned use & Dynamic Population Results: Residual Activity a res ij (t) =a norm (t) ā ij norm (t) Inverse relationship between RESIDENTIAL and COMMERCIAL Hour Residential Industrial Parks Other Weekends are quantitatively different than weekdays. 10

11 Zoned use & Dynamic Population static population Abs. Act phone use 1 0 movement Hour Parks Other 11

12 Absolute Activity Results: Low Activity High Activity 12

13 Inferring Land Use Methods - Supervised Learning: Random Decision Tree Θ 1 1 <y 1 h(x, Θ k ) = c Θ 1 2 <y 2 Θ 1 3 <y 3 c 1 c 2 c 3 c 4 13

14 Inferring Land Use Methods - Supervised Learning: Random Forest Random Forest: (Breiman 2001) 1.1. A random forest i {h(x, k ), k = 1,...} ectors and each tree casts x is a T 1vector Θ k is a m 1vector,(m<T) Θ 1 1 <y Θ k 1 <z 1 Θ 1 2 <y 2 Θ 1 3 <y 3 Θ k 2 <z 2 Θ k 3 <z 3 c 1 c 2 c 3 c 4 c 1 c 2 c 3 c 4 14

15 Inferring Land Use Methods - Supervised Learning: Random Forest Random Forest: (Breiman 2001) 1.1. A random forest i {h(x, k ), k = 1,...} ectors and each tree casts x is a T 1vector Θ k is a m 1vector,(m<T) Θ 1 1 <y Θ k 1 <z 1 Θ 1 2 <y 2 Θ 1 3 <y 3 Θ k 2 <z 2 Θ k 3 <z 3 c 1 c 2 c 3 c 4 c 1 c 3 c 4 c 2 {h(x, k ) } = } {c2c2 c3 c2 c1 c2 c1 c4 ectors and c ^ = mode( {h(x, k ) }) perf = N ^ i I(c i = j) N ectors and 15

16 Results: Inferring Land Use Confusion Matrix Thresh Total Accuracy Weights Res Com Ind Prk Oth ( ) Res Com Ind Prk Oth Share: Res Com Ind Prk Oth 16

17 Inferring Land Use Results: Confusion Matrix Thresh Total Accuracy Weights Res Com Ind Prk Oth ( ) Res Com Ind Prk Oth Share: = 1 Res Com Ind Prk Oth

Results: Inferring Land Use Confusion Matrix Thresh Total Accuracy Weights Res Prk Ind Con Oth 500 0.396 ( N/A 0.30 0.30 0.20 0.

18 Results: Inferring Land Use Confusion Matrix Thresh Total Accuracy Weights Res Prk Ind Con Oth ( N/A ) N/A N/A N/A N/A N/A N/A N/A Res Com Ind Prk Oth N/A Share: N/A

19 Inferring Land Use Classification Group I Group II Group III Error Analysis Cells correctly predicted to be a Cells of a given use incorrectly Cells of a different use incorrectly predicted to be a given Residential I II III Commercial I II III 19

Robustness Across Cities Cross Tabulation Thresh Avg. Error Cutoffs Res Com Ind Par Oth 100 0.468 ( 0.50 0.12 0.12 0.12 0.12) 0.66 0.09 0.08 0.05 0.12 0.53 0.

20 Robustness Across Cities Cross Tabulation Thresh Avg. Error Cutoffs Res Com Ind Par Oth ( ) Res Com Ind Prk Oth Share:

21 Robustness Across Cities 80 City wide Time Series Chicago LBS Data Abs. Act Norm. Act Res. Act Hour 21

22 Inferring Land Use Summary: Introduced a way to reconcile and normalize traditional and new data sources Used temporal activity patterns to identify land uses Performed error analysis to determine why classifications failed Similar results obtained for different cities Implications: Mobile phone activity can be radically different depending on zoned land use. Traditional zoning maps should be augmented to account for real-time use. Dynamic land use patterns seem relatively consistent despite differences in history and regulatory classification 22

23 Future Contributions Use more sophisticated feature spaces for zoning classification. Introduce new data sources. Measure urban system dynamics in multiple cities around the globe to identify similarities and differences in behavior. Apply normalization metrics to other data sources. Combine our understanding of temporal land use with measurements of human mobility to develop better models of intra-city travel behavior. 23

24 Thank You! Questions? 24

25 Errors 25

26 Correlation with Static Population 26

27 Group I Group II Group III Classification Error Analysis Cells correctly predicted to be a Cells of a given use incorrectly Cells of a different use incorrectly predicted to be a given Residential I II III Commercial Industrial I II III I II III Parks Other I II III I II III 27

Chapter 5 FUNCTIONAL CLASSIFICATION

Chapter 5 FUNCTIONAL CLASSIFICATION Functional classification is a system by which streets and roadways may be distinguished by types according to their function within the entire transportation network.