LSTM with Matrix Factorization for Road Speed Prediction.

Road speed prediction is a key point of Intelligent Transport System. Plenty of work have proved the effectiveness and efficiency of neural network in forecasting freeway velocity. However, the missing values are obstacles when applying the widely used trajectory data to neural network. In trajectory data, most roads may not be covered by enough trajectories in a short time. Due to highly sparsity, it will bring extra cost if we first fill missing data then perform training. To solve this issue, we propose a collaborative model that combines LSTM neural network with matrix factorization to reduce sparsity and make prediction simultaneously. We conduct experiments with a sufficient amount of trajectories and the results show that our model outperforms cascaded methods in both MAE and RMSE.