City:

Boston, Massachusetts, USA 

Organization:

Language Technologies Institute at Carnegie Mellon University in Pittsburgh, Pennsylvania, USA 

Project Start 

Date: 

Project End Date:

2015 (Published) 

Reference:

Oh, Jean. (2015). Finding Main Streets: Applying Machine Learning to Urban Design Planning 

Problem:

An integral component of the urban design discipline focuses on the collection, integration and typological analysis (classification) of various urban features. This task (also referred to as Site inventory and Analysis) is often a time-consuming and expensive undertaking. 

This study tackles these challenges by developing an active learning model that analyses urban typologies to help classify a certain type of urban setting, particularly, Main Streets. 

Ultimately: Applying machine learning techniques to determine if a certain district within a city should be considered a Main Street or not. 

Technical Solution:

Supervised (Predictive) Machine Learning 

·       Target concept: Main Streets, are defined in districts which are composed of several commercial buildings (through industry consultation, it was determined that Main Streets typically appear in high-density and commercially zoned areas of the city) 

Preprocessing the data into set of tuples: 

·       A data preprocessor was applied to cluster the GIS buildings data into a set of candidate districts (Since Main Streets are typically organized as commercial districts, the GIS Building Layer was clustered by commercial buildings within a certain proximity boundary). Small clusters that had >10 commercial buildings were filtered out at this phase. This narrowed the total 180,000 data points (from the GIS buildings and parcels data) to 80 district candidates for Main Streets. 

The following Binary Classifiers were applied: 

·       k-Nearest Neighbor (k-NN) classifier

·       Naïve Bayes classifier

·       Decision Trees classifier
[Selected model due to ease of visualization and comprehension]

·       Support Vector Machine (SVM) classifier
[Best Performer]

Datasets Used:

Training Data: 

·       Dataset 1: City of Boston Main Streets Data Inventory (19 Districts were pre-identified as Main Streets by field experts) 

Testing Data: 

·       Dataset 2: City of Boston GIS Building Layer

·       Dataset 3: City of Boston GIS Parcels Layer 

Outcome:

The Experiment results had two major outcomes/contributions:  

·       Architectural typology problems can be modelled as a classification problem (based on GIS data and the decision tree classifier, the model was able to label Main Streets / not Main Streets across City of Boston) 

·       Learning Speed of the Decision Tree Model can be accelerated with more feature data sets (More ways to identify Main Streets besides clusters of Commercial Buildings) 

Given the limited access to data, the results were stated as reasonably good and convincing 

Issues that arose:

·       Integration of Heterogeneous information / data was difficult 

·       Industry designers use a lot more features to help classify streets than the ones used for this experiment 

Status:

On going: efforts to create a new user interface with various learning components as an “Urban Planner Assistant System”. Including data analysis and inference capabilities through machine learning techniques. 

The research team aims to make this a more comprehensive experiment, by trying this procedure in other cities that have a different City form/urban fabric than the City of Boston 

Entered by:

Sabrina Samin, sabrina.samin@mail.utoronto.ca