Changxing Shao

LSTM (Long-short Term Memory) is an effective method for trajectory prediction. However, it needs to rely on the state value of the previous unit when calculating the state value of neurons in the hidden layer, which results in too long training time and prediction time. To solve this problem, we propose Fast Trajectory Prediction method with Attention enhanced SRU (FTP-AS). Firstly, we devise an SRU (Simple Recurrent Units) based trajectory prediction method. It removes the dependencies on the hidden layer state at the previous moment, and enables the model to perform better parallel calculation, speeding up model training and prediction. However, each unit of the SRU calculates the state value at each moment independently, ignoring the timing relationship between the track points and leading to accuracy decrease. Secondly, we develop the attention mechanism to enhance SRU. The influence weight for selective learning is gained by calculating the matching degree of the hidden layer state value at each moment to improve the accuracy of the prediction. Finally, experimental results on MTA bus data set and Porto taxi data set showed that FTP-AS was 3.4 times faster and about 1.7% more accurate than the traditional LSTM method.

How to predict spatiotemporal activity from geo-tagged social media is an urgent problem. Existing methods don’t make full use of spatiotemporal information and text sequence features. In view of above problem, we design a Fast Lightweight Spatiotemporal Activity Prediction method(FLSAP) based on Gated Recurrent Unit(GRU) neural network. While GRU structure can extract text sequence features, the model takes up a lot of space due to the numerous parameters. At the same time, due to the long sequence in the text, the convergence speed of GRU is slow. So, we design a novel GRU neuron, GRU with Tiny and Skip(GTS), which can quickly generate a lightweight model with higher accuracy. In GTS, we add a scalar weighted residual connection to stabilize the training. Furthermore, we extend the residual connection to a gate by reusing the parameter matrices to compress the model size. At last, in order to make the model converge faster, we add a binary gate, which determine whether to skip the current state update. According to the experimental results, compared with ReActzhang2017react in the spatiotemporal activity prediction task, FLSAP improves the accuracy by 3.3%, reduces the model space by 98.79% and accelerates 74.4% of convergence speed.