Top Banner
Learning from Main Learning from Main Streets: Streets: A Machine Learning Approach Identifying Neighbourhood A Machine Learning Approach Identifying Neighbourhood Commercial Districts Commercial Districts DDSS 2006 International Conference on Design & Decision Support Systems in Architecture and Urban Planning Jean Oh Stephen F. Smith School of Computer Science, Carnegie Mellon University Jie-Eun Hwang Graduate School of Design, Harvard University Kimberle Koile Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology
22

DDSS 2006

Jan 03, 2016

Download

Documents

callum-pace

Learning from Main Streets: A Machine Learning Approach Identifying Neighbourhood Commercial Districts. DDSS 2006 International Conference on Design & Decision Support Systems in Architecture and Urban Planning Jean Oh Stephen F. Smith School of Computer Science, Carnegie Mellon University - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: DDSS 2006

Learning from Main Learning from Main Streets:Streets:A Machine Learning Approach Identifying Neighbourhood Commercial A Machine Learning Approach Identifying Neighbourhood Commercial DistrictsDistricts

DDSS 2006International Conference on Design & Decision Support Systems in Architecture and Urban Planning

Jean OhStephen F. SmithSchool of Computer Science, Carnegie Mellon University

Jie-Eun HwangGraduate School of Design, Harvard University

Kimberle KoileComputer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology

Page 2: DDSS 2006

MotivationMotivation

• Data analysis• Physical, social, &

economic constraints• Interdisciplinary “multiple

views” problem

• Information retrieval, machine learning

• Constraint reasoning• Distributed AI (Multiagent

systems)

Design A.I.

Page 3: DDSS 2006

Urban typologyUrban typology

• Formalize urban components for better communication among diverse interest groups

• Classification• Human experts sanctuary

Page 4: DDSS 2006

Intelligent Urban Design Intelligent Urban Design AssistantAssistant

• GIS and beyond

• Efficient urban typology: machine learning E.g., Main Streets experiment

• Learning in a distributed environment

Page 5: DDSS 2006

ARTISTS: ARTISTS: ArtArtereriial al SStreets treets TTowards owards SSustainabilityustainability

(Svensson et al. 2004)(Svensson et al. 2004)

• Human experts• Duration: 3 years (~2004)• Budget: 2.2 million euros• Classified 40 streets in 9 countries into 5 categories

Page 6: DDSS 2006

ARTISTS typology of arterial streets

Low Intensity Street

Narrow Inactive Old Street

Shopping Street

Metropolitan Arterial Suburban Residential Arterial

Page 7: DDSS 2006

Main StreetsMain Streets

• the generic street name of the primary retail street of an urban area, especially a village or town, in many parts of the world. It is usually a focal point for shops and retailers in the city centre, and is most often used in reference to retailing

Historiography of Townscape Icon of Townscape Design Process of Townscape

Page 8: DDSS 2006

Finding Main StreetsFinding Main Streets

• Why Main Streets Matter: A series of individual structures become townscape. Diverse participants have various perspectives on community

development. Historic preservation brings controversial issues.

Need Heuristic Process to interpret existing context!

• Information sources: GIS data • Criteria (features)

Building/parcels structural data Land use Business types, etc.

A machine learning approach!

Page 9: DDSS 2006

Machine Learning ApproachMachine Learning Approach

• Clustering: unsupervised learning• Classification: supervised learning• Active Learning: fast learning

Page 10: DDSS 2006

Finding Main StreetsFinding Main Streets

Buildings, parcels, tuple dataGIS

ParcelBuilding

Data export

Page 11: DDSS 2006

Feature space modeling Feature space modeling (survey)(survey)

Public-ness

BuiltForm Function

Use Patterns

Streetscape Building

Massing

Lot(Parcels)

Frontage Sign

Yard Entrance Front Transparency

Type of Signage

AbstractClass

Semantic Class

Feature From DB

Feature By User

Feature Intangible

Popularity

Business Type

Height, Area,

Periphery,Distance,

.

.

ArchitecturalStyle

Types of User

Groups

Types of Activities

Population of People

Quality of Maintenance /

Service

Awareness of Content

Distance,Area,

Vegetation,..

Num of Door,Stair Size,

.

.

Num of Window,

Dimension of Windows,Material of Windows

Location,Size,

Material...

Visibility

Bold Line : User Annotatable

Legends

Page 12: DDSS 2006

Finding Main StreetsFinding Main StreetsBuildings data

Parcel

GIS

Building

Form candidate districts

Unsupervised Learning: Clustering

Data export

Page 13: DDSS 2006

Clustering (single linkage)Clustering (single linkage)

What defines “distance” between two data points?

Page 14: DDSS 2006

Main Street Candidates Main Street Candidates (Boston)(Boston)

90,649 buildings99,897 parcels

4,049 commercial

76 candidate districts

Page 15: DDSS 2006

Finding Main StreetsFinding Main StreetsBuildings data

Main Street Prediction

candidate districts

Parcel

GIS

Building

Form candidate districts

Unsupervised Learning: Clustering

Supervised Learning: Classification

Data export

Page 16: DDSS 2006

Classification Classification

Page 17: DDSS 2006

Active Learning with SVMActive Learning with SVM(Support Vector Machine)(Support Vector Machine)

Support vectors

?

Page 18: DDSS 2006

Finding Main StreetsFinding Main Streets

Buildings data

Main Street Prediction

candidate districts

Parcel

GIS

Building

Form candidate districts

Initial train data

Next district to be labeled

predictions

Active Learner

Unsupervised Learning: Clustering

Supervised Learning: Classification

Active Learning with SVM

Data export

Page 19: DDSS 2006

Evaluation MetricsEvaluation Metrics

• n : total # of examples • m: total # of Main Streets in Boston• a: # of examples classified as Main Streets• c: # of correct Main Streets in the answers

Precision: p = c/aRecall: r = c/mF1 = 2pr / (p + r)

Page 20: DDSS 2006

Results

Precision Recall F1

LOOCV 0.842 0.762 0.800

Leave-One-Out-Cross-Validation

Page 21: DDSS 2006

Conclusion and future Conclusion and future directionsdirections

• Urban design decision support system can benefit from machine learning approaches.

• The need for such support has been underscored after series of failures of recent post-disaster management.

• Comparison with morphological approaches• Learning in a multiagent environment

Page 22: DDSS 2006

Thank you!Thank you!