Grids-versus-Graphs
The widely employed ConvLSTM for modeling spatio-temporal data can model only data in fixed-sized partitions. The standard convolutional filters cannot be applied on variable-sized partitions. The search for a model that works well with irregular spaced partitions made us delve into Graph-based LSTM models. A comparison of ConvLSTM and GraphLSTM reveals the competitive performance of GraphLSTM, at a lower computational complexity, across three real-world large-scale taxi demand-supply data sets, with different performance metrics.
For graph-based modeling, we borrow heavily from Cui, Zhiyong, et al. "Traffic graph convolutional recurrent neural network: A deep learning framework for network-scale traffic learning and forecasting." and their code at https://github.com/zhiyongc/Graph_Convolutional_LSTM
For ConvLSTM and LSTM baselines, we use the standard Keras libraries.