Transportation Research Board 87th Annual MeetingTransportation Research Board · 2008 · 30 citations · 0 references
ComputingData ScienceMachine LearningPredictive AnalyticsTransportationIntelligent Transportation SystemsPredictive ModelingShort-term ForecastingSupport Vector MachinesTime Series ForecastingTime Series PredictionMachine Learning PotentialsMachine Learning-based ForecastingNearest Neighbor RetrievalShort-term Traffic PredictionSupport Vector RegressionTime-varying Coefficient ModelsTrainingTravel Time Prediction
This paper presents an investigation into application potentials of Support Vector Regression (SVR), an advanced technique in machine learning, for time-series travel time prediction. One-month traffic data on a segment of the Pan-Island Expressway in Singapore is used for the model training and testing. Two baseline predictors, including the historical mean predictor and the time-varying coefficient model, are employed in the comparative study. The results show that the SVR significantly outperformed the baseline predictors in both normal and congested traffic conditions over a wide range of prediction headways. To improve the prediction performances, an empirical Nearest Neighbor method is introduced to retrieve patterns closest to the test vector for SVR training. The results show that in SVR-based prediction, the Nearest Neighbor method allows a substantial reduction of training size to accelerate training but to maintain the prediction accuracy, and in time-varying coefficient model the prediction accuracy is significantly improved if reference data are retrieved from the Nearest Neighbor method than from historical mean or median.