Correlating Extreme Weather Conditions With Road Traffic Safety: A Unified Latent Space Model

The presence of extreme weather conditions is known to expose drivers to a higher risk to incur in road accidents. Quantifying the correlation between adverse weather conditions and road traffic safety is useful for several reasons such as planning preventive actions, managing vehicle fleets, and configuring alerting systems. However, since the risk of road accidents occurrences within a specific spatial region is influenced by several factors other than the weather conditions, quantifying the actual impact of adverse weather phenomena regardless of the effect of weather-unrelated conditions can be challenging. To tackle the aforesaid issue, this paper proposes to adopt a unified latent space model based on time series embeddings. Firstly, it encodes a subset of historical series reporting weather-related accident occurrences in specific risky areas into the high-dimensional vector representation. It also encodes the weather element measurements acquired by meteorological stations spread over the analyzed area. Then, to estimate the risk level of each region within the same spatial context it seeks the temporal risk patterns that are most similar to those observed in risky areas. The experiments carried out in a real case study confirm the applicability of the proposed approach.