IR term project of Jean Oh

Basic Information

• Project Title: Finding Main Streets: Applying Machine Learning to Urban Design
• Name: Jean Oh (email: jeanoh _at_ cs cmu  edu)
• Presentation Date: 8/30 (tentative)
• Demo Date: 8/30 (tentative)

A PDF version of project report and source code are available upon request (jean oh _at_ cs. cmu. edu).

* This project is a joint work with Jie-Eun Hwang at Graduate School of Design, Harvard University, Cambridge, Boston.

Contents

• Abstract
• Proposal and Timelines
• System Description
• Experiments
• Results
• Demo

Abstract

In the urban design realm careful consideration of connecting architectural form and socioeconomic function is a compelling issue. City planners and architects spend a significant amount of their time on collecting and integrating data from various information sources. Rapid growth of Geographic Information Systems (GIS) made it possible to map and display laboriously collected data, but these tools are limited by lack of sophisticated data analysis and inference capability. In this project we explored possibilities of how A.I. techniques can boost the performance of urban design planning by providing large scale data analysis and inference capability. In addition to common benefits of automated process such as speed and labor cost, statistical analysis can also provide theoretical justification for designers which is not typically available in the case of manual efforts. As a proof of concept experiment we implemented an application of active learning that identifies a certain type of urban setting, Main Streets, based on their complicated spatial and semantic relationships over building geometry. The preliminary results show that active learning algorithm can effectively learn a classifier with relatively small number of training examples.

Proposal and Timelines

• GIS (Data from City of Boston, publicly available GIS data source)
• Main Streets (Semi-structured web pages)

Related work

• ARTISTS (Arterial Steets Towards Sustainability): The ARTISTS project in Europe is a good example that shows the current trend of architectural typology and also shows how laborious and expensive this process is. The main goal of ARTISTS was to identify types of streets in order to provide better insights to urban planners and economists. This 2.2 billion euros budget project involved 17 European countries and took three years to classify five categories of streets. Their main effort was made to statistically analyze the characteristics of street functions and draw a two-dimensional classification table. This project was completed by human experts. Their experimental results include how they classified 48 streets into 5 categories based on their two-dimensional classification table.

Literature Survey

Feature Selection

• A comparative study on feature selection in text categorization, Yiming Yang, Jan O. Pedersen, ICML 1997[ps]

Clustering

• Learning Spectral Clustering, Francis R. Bach and Michael I. Jordan, NIPS 2003 [pdf]
• Spectral Clustering tutorial by Chris Ding

Active Learning

• Active Learning Using Pre-clustering, Hieu T. Nguyen and Arnold Smeulders, ICML 2004 [pdf]

Semi-supervised learning (using unlabeled data to guide supervised learning)

• Co-Trainig: Combining Labeled and Unlabeled Data with Co-Training. Avrim Blum and Tom Mitchell. Proceedings of the 11th Annual Conference on Computational Learning Theory, pages 92--100, 1998.
• Schuurmans, D. and Southey, F. (2002) Metric-based methods for adaptive model selection and regularization. Machine Learning 48(1-3): 51-84. Special Issue on New Methods for Model Selection and Model Combination [ps.gz]
• Text Classification from Labeled and Unlabeled Documents using EM. Kamal Nigam, Andrew McCallum, Sebastian Thrun and Tom Mitchell. Machine Learning, 39(2/3). pp. 103-134. 2000 [PS]

Spatial data mining from GIS

• SPIRIT (Spatially-Aware Information Retrieval on the Internet)

Inference network from relational data

• Learning Statistical Models from Relational Data, L. Getoor. Ph.D. Thesis, Stanford University, December, 2001.

Experiments

As a proof-of-principle experiment we built an interactive Main Street Finder that utilizes active learning technique to classify certain urban settings. We used main streets data from the city of Boston, Massachusetts.
Figure 1 Main Streets in Boston, Massachusetts

Results

We explored possibilities of interesting A.I. research in urban planning domain. Since the urban planning community is conservative towards computational assistance our primary goal was to implement a simple example prototype to showcase potentials of applying machine learning and A.I. techniques. In our preliminary experiment of finding Main Streets, the results has two major contributions. First, we showed that architectural typology problem can be modeled as a classification problem except the fact that classification is more generalized paradigm. Second, using active learning our system can cleverly choose better samples to label, outperforming random selection model significantly as shown in Figure 2.

Figure 2 Random Selection vs. Active Learning Results