Multi-modal Sequence to Sequence Learning with Content Attention for Hotspot Traffic Speed Prediction

Abstract

Traffic speed prediction is a crucial and fundamental task of the intelligent transportation systems (ITS). Due to the dynamic and non-linear nature of the traffic, this task is difficult. Nonetheless, the collection of crowd map queries data brings new ways to solve this problem. Generally speaking, in a short period of time, a large amount of crowd map queries aiming at the same destination may lead to traffic congestion. For instance, large queries for Family Restaurant during the dinner time lead to traffic jams around it. However, traffic speed prediction with crowd map queries is challenging due to the complexity and scale of the map queries, as well as their modalities. To bridge the gap, we propose Multi-Seq2Seq-Att for hotspot traffic speed prediction. Multi-Seq2Seq-Att is a multi-modal sequence learning model that deals with two sequences in different modalities, namely, the query sequence and the traffic speed sequence. The main idea of Multi-Seq2Seq-Att is to learn to fuse the multi-modal sequence with content attention. With this method, Multi-Seq2Seq-Att addresses the modality gap between queries and the traffic speed. Experiments on real-world datasets from Baidu Map demonstrates a 24% relative boost over other state-of-the-art methods.