Abstract:Traffic accidents pose a significant risk to human health and property safety. Therefore, to prevent traffic accidents, predicting their risks has garnered growing interest. We argue that a desired prediction solution should demonstrate resilience to the complexity of traffic accidents. In particular, it should adequately consider the regional background, accurately capture both spatial proximity and semantic similarity, and effectively address the sparsity of traffic accidents. However, these factors are often overlooked or difficult to incorporate. In this paper, we propose a novel multi-granularity hierarchical spatio-temporal network. Initially, we innovate by incorporating remote sensing data, facilitating the creation of hierarchical multi-granularity structure and the comprehension of regional background. We construct multiple high-level risk prediction tasks to enhance model's ability to cope with sparsity. Subsequently, to capture both spatial proximity and semantic similarity, region feature and multi-view graph undergo encoding processes to distill effective representations. Additionally, we propose message passing and adaptive temporal attention module that bridges different granularities and dynamically captures time correlations inherent in traffic accident patterns. At last, a multivariate hierarchical loss function is devised considering the complexity of the prediction purpose. Extensive experiments on two real datasets verify the superiority of our model against the state-of-the-art methods.
What problem does this paper attempt to address?
This paper attempts to solve several key problems in traffic accident risk prediction, specifically including:
1. **Regionality**: Traffic accidents are affected by many complex factors, and it is difficult to analyze how these factors affect traffic accidents. Therefore, capturing the regional context is crucial for understanding the underlying mechanisms of traffic accidents. Different regions may have similar road structures and weather conditions, but the regional contexts (such as parking lots, lakes, small parks, and dense residential areas) are completely different, which leads to significant differences in traffic accident patterns.
2. **Proximity and Similarity**: Traffic accidents show a dual correlation, involving both spatial proximity and semantic similarity. This means that traffic accidents occurring in adjacent areas or areas with similar semantic properties tend to be highly similar. For example, roads in adjacent areas are connected, and areas with similar point - of - interest (POI) distributions may have similar accident patterns. Effectively capturing these two correlations is a major challenge.
3. **Sparsity**: There is a zero - inflation problem in traffic accident data, that is, most areas do not have traffic accidents most of the time. If this sparsity cannot be properly handled, the model may generate too many zero predictions, making the prediction results meaningless.
To address these problems, the authors propose a new Multi - Granularity Hierarchical Spatio - Temporal Network (MGHSTN), aiming to comprehensively utilize spatio - temporal features for traffic accident risk prediction. Specific improvement measures include:
- **Introducing remote sensing images**: Enhance the understanding of the regional context through remote sensing images and create multi - level regional features and multi - view graph structures.
- **Constructing multi - view semantic similarity graphs**: Combine information such as road structures, risk features, and point - of - interest distributions to construct semantic similarity graphs of multiple views to capture semantic similarity.
- **Designing an adaptive temporal attention module**: By introducing an adaptive temporal attention mechanism, dynamically capture time - dependent relationships, thereby better handling sparsity and time - series correlations.
- **Multi - level embedding fusion mechanism**: Through a multi - level embedding fusion module, fully utilize the information transfer between different granularity levels to improve prediction accuracy.
In summary, this paper aims to comprehensively solve the problems of regionality, proximity, similarity, and sparsity in traffic accident risk prediction through the MGHSTN model, thereby providing more accurate and robust prediction results.