面向交通事故预测的时空多模态点过程

交通事故预测对于构建智慧城市具有重要意义。然而发生在连续时间域上的交通事故数据同时包含具有不同语义特征的时间、空间模态信息，且这两种模态的不确定性存在差异，因此传统的序列建模方式无法全面描述交通事故的时空相关性，很难实现准确的交通事故预测，对此提出了一种面向交通事故预测的时空多模态点过程模型MSTPP。该模型设计了一种具有双解码器的seq2seq框架。在编码器中提出了衰减感知长短期记忆网络DLSTM用于编码在连续时间域中的交通事故事件序列，有效地融合不同模态信息以及建模事件序列的异步性。在解码阶段，使用两个特殊设计的解码器去处理模态间差异性。在两个真实的交通事故数据集上的实验结果表明，MSTPP在预测下一个交通事故发生的时间和区域任务上相比于其他基准模型具有最优的预测能力。

Multimodal spatial-temporal point processes for traffic accident event prediction

Traffic accident event prediction is of great importance to build intelligent transportation systems. However, traffic accident event data occurring in the continuous time domain contains temporal and spatial modal information with different semantic characteristics and different uncertainty, so the traditional sequence models cannot fully describe the spatial-temporal correlation of traffic accident events, and it is difficult to achieve accurate traffic accident prediction. So this paper proposed a multimodal spatial-temporal point process(MSTPP) model. And the model designed a seq2seq framework with dual decoders. It proposed decay-aware long short-term memory networks(DLSTM) in the encoder for encoding traffic accident event sequences in the continuous time domain, effectively fusing different modal information and modelling the asynchronicity of event sequences. In the decoding stage, it used two specially designed decoders to handle the difference between the two modalities. Extensive experiments on two real-world datasets demonstrate the superiority of MSTPP against the state-of-the-art baseline methods with regard to both the next accident happening time prediction and region prediction tasks.