Predicting Travel Times for the M4 Highway in Sidney: How We Won Kaggle.com's First $10,000 Data Mining Challenge

• Activity Number: 104
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: General Methodology
• Author(s): José Pablo González-Brenes*+
• Companies: Carnegie Mellon University
• Keywords: Forecasting travel time ; Random Forest ; Data mining competition

Abstract

The New South Wales Roads and Traffic Authority collected over 2 years worth of historical data on road use between 2008 and 2010 and made it available for a data mining competition on kaggle.com. The challenge consisted of forecasting travel time on the M4 highway in different prediction horizons - ranging from a few minutes in the future, to up to 24 hours ahead. Contestants' submissions were evaluated on how accurately travel time was predicted on a held-out dataset. We describe how we built the statistical model that outperformed over 350 other teams.

We model highway travel time using a non-parametric algorithm called Ensemble of Decision Trees. This approach is able to efficiently capture non-linear interactions between regressors. We compare Ensemble of Decision Trees with auto-regressive models and describe some of the difficulties we encounter, such as the non-i.i.d nature of the held-out set.