Accurate prediction of traffic flow is of great significance for alleviating urban traffic congestions. Most previous studies used historical traffic data, in which only one model or algorithm was adopted by the whole prediction space and the differences in various regions were ignored. In this context, based on time and space heterogeneity, a Classification and Regression Trees-K-Nearest Neighbor (CART-KNN) Hybrid Prediction model was proposed to predict short-term taxi demand. Firstly, a concentric partitioning method was applied to divide the test area into discrete small areas according to its boarding density level. Then the CART model was used to divide the dataset of each area according to its temporal characteristics, and KNN was established for each subset by using the corresponding boarding density data to estimate the parameters of the KNN model. Finally, the proposed method was tested on the New York City Taxi and Limousine Commission (TLC) data, and the traditional KNN model, backpropagation (BP) neural network, long-short term memory model (LSTM) were used to compare with the proposed CART-KNN model. The selected models were used to predict the demand for taxis in New York City, and the Kriging Interpolation was used to obtain all the regional predictions. From the results, it can be suggested that the proposed CART-KNN model performed better than other general models by showing smaller mean absolute percentage error (MAPE) and root mean square error (RMSE) value. The improvement of prediction accuracy of CART-KNN model is helpful to understand the regional demand pattern to partition the boarding density data from the time and space dimensions. The partition method can be extended into many models using traffic data.
Urban Traffic Data Imputation With Detrending and Tensor Decomposition
