GENERALIZATION ABILITY IN PREDICTING ROAD OPENINGS FROM CONSTRUCTION DATA
Wataru Kobayashi1, Ryosuke Shibasaki2 and Yoshihide Sekimoto3
1) MEng, Prof., Center for Research and Collaboration, Tokyo Denki University, Tokyo, Japan
2) Ph.D., Prof., Center for Spatial Information Science, University of Tokyo, Tokyo, Japan.
3) Ph.D., Assoc. Prof., Institute of Industrial Science, University of Tokyo, Tokyo, Japan.
Abstract: We have been studying methods for predicting road openings in order to make updating road maps and navigation data more efficient. To predict road openings, we used supervised machine learning and construction data consisting of construction locations, construction completion times, and construction items. Training data in supervised machine learning creates a learning model, and its predictive accuracy decreases when the training data and the learning model is not appropriate to the target (test data). This paper studies the ability to generalize—that is, the ability to predict road openings from new (unknown) construction data with already existing training data. Our experiment shows that learning models from three road administrators revealed parts that were both similar and different. Predictions are considered accurate if they predict openings that actually open within six months of the prediction. Given this condition, the three models can accurately predict more than 80% of openings for unknown test data.
Keywords: Supervised Machine Learning, Data Mining, Generalization Ability, Construction data, Road Map
Wataru Kobayashi, Ryosuke Shibasaki, and Yoshihide Sekimoto. “GENERALIZATION ABILITY IN PREDICTING ROAD OPENINGS FROM CONSTRUCTION DATA.” In Proceedings of International Conference on Civil and Building Engineering Informatics (ICCBEI 2015), 94. Tokyo, Japan, 2015.