PROJECT TITLE :

Spatio-Temporal Meta Learning for Urban Traffic Prediction

ABSTRACT:

It is very difficult to predict urban traffic because of three factors: 1) the complex spatio-temporal correlations of urban traffic, which include spatial correlations between locations along with temporal correlations among timestamps; 2) the spatial diversity of such spatio-temporal correlations, which varies from location to location and depends on the surrounding geographical information, such as points of interest; and 3) the importance of accurately predicting urban traffic to intelligent transportation systems and public safety. To address these issues, we came up with the idea of developing a model that is based on deep metalearning and is given the name ST-MetaNet +. This model is intended to predict traffic in all locations simultaneously. A sequence-to-sequence architecture is used by ST-MetaNet +. This architecture consists of an encoder to learn information about the past and a decoder to make predictions in a step-by-step fashion. Specifically, the encoder and the decoder share the same network structure, which consists of meta graph attention networks and meta recurrent neural networks, in order to capture a wide variety of spatial and temporal correlations, respectively. In addition, the embeddings of geo-graph attributes and the traffic context that is learned from dynamic traffic states are used to generate the weights, also known as parameters, of meta graph attention networks and meta recurrent neural networks. Extensive experiments based on three real-world datasets were carried out in order to demonstrate that ST-MetaNet + is more effective than a number of other methods that are considered to be state-of-the-art.