2019/02/18 by Neema Davis, Davis, Neema, Gaurav Raina +3
Engineering · Social Sciences · #Traffic Prediction and Management Techniques #Transportation and Mobility Innovations #Transportation Planning and Optimization
paper · pdf · doi:10.48550/arxiv.1902.06515
Accurate taxi demand-supply forecasting is a challenging application of ITS\n(Intelligent Transportation Systems), due to the complex spatial and temporal\npatterns. We investigate the impact of different spatial partitioning\ntechniques on the prediction performance of an LSTM (Long Short-Term Memory)\nnetwork, in the context of taxi demand-supply forecasting. We consider two\ntessellation schemes: (i) the variable-sized Voronoi tessellation, and (ii) the\nfixed-sized Geohash tessellation. While the widely employed ConvLSTM\n(Convolutional LSTM) can model fixed-sized Geohash partitions, the standard\nconvolutional filters cannot be applied on the variable-sized Voronoi\npartitions. To explore the Voronoi tessellation scheme, we propose the use of\nGraphLSTM (Graph-based LSTM), by representing the Voronoi spatial partitions as\nnodes on an arbitrarily structured graph. The GraphLSTM offers competitive\nperformance against ConvLSTM, at lower computational complexity, across three\nreal-world large-scale taxi demand-supply data sets, with different performance\nmetrics. To ensure superior performance across diverse settings, a HEDGE based\nensemble learning algorithm is applied over the ConvLSTM and the GraphLSTM\nnetworks.\n