Reliable Event Detection via Multiple Edge Computing on Streaming Traffic Social Data

Abstract

Traffic event detection plays an essential role in flexible decision-making for sensor-cloud system(SCS). Since a social traffic event is often described by multiple social media texts, it is significant to perform traffic event detection on streaming social media texts. However, these social media texts with limited semantic features fail to describe a large amount of traffic event categories and a small number of samples per traffic event category, which are difficult to solve with conventional text classification methods and reduce the reliability in SCS. In this paper, we propose a reliable and streaming clustering algorithm for streaming traffic event detection via multiple edge computing. First, we combine traffic-related knowledge information to extract various types of elements from social media texts, and accordingly construct a traffic event-based heterogeneous information network (HIN) and proceed to calculate event similarity between social media texts through meta-path weights. Then, we utilize graph neural networks to perform semi-supervised learning on HIN to obtain the optimal meta-path weights. We also develop Binary Sample Graph Convolutional Neural Network (BS-GCN) and Binary Sample Graph Attention Network (BS-GAT) to improve the reliability of graph neural network models based on the characteristics of traffic event detection and design an incremental clustering algorithm based on event similarity to implement streaming social traffic event detection. We conduct experiments on social media text datasets describing various traffic events in cities such as Beijing and develop related social traffic event detection systems. The results indicate that our model can better implement streaming social traffic event detection, and is superior to most text classification methods.