Stacking LSTMs to Extract Features in Two-Dimensional Data for Prediction Tasks on Travel Time and Crimes Frequency

Data-based methods have some advantages over other methods for prediction tasks, with domain knowledge is not required. In these methods, Long-short Term Memory has achieved tremendous success in recent years, because its capable of feature extraction of time series. However, there are fewer studies on using Long-short Term Memory to extract the features in spatial dimension. In our work, we attempt to stack Long-short Term Memorys to extract the spatial dimensional features while completing feature extraction of time series for prediction tasks. An multilayer stacking method with multiway inputs and an grid mode stacking method were proposed to extract the features in two dimensions. In the grid mode stacking method, an memory block was proposed to merge the gated cell vectors and each Long-short Term Memory is for one embedding vector separately. Based on two public datasets, we evaluated the proposed methods on travle time prediction and crime prediction. Travel time dataset is provided by Highways England, and crime dataset is provided by the Home Office. Experimental results on the datasets demonstrate that the proposed methods perform well compared to the baseline methods.