A Knowledge-Driven Memory System for Traffic Flow Prediction

Binwu Wang, Yudong Zhang, Pengkun Wang, Xu Wang, LEI BAI, Yang Wang

Abstract: Traffic flow prediction is critical for intelligent transportation systems. Recent studies indicate that performance improvement by designing new models is becoming marginal. Instead, we argue that the improvement can be achieved by using traffic-related facts or laws, which is termed exogenous knowledge. To this end, we propose a knowledge-driven memory system that can be seamlessly integrated into GCN-based traffic forecasting models. Specifically, the memory system includes three components: access interface, memory module, and feedback interface. The access interface based on the attention mechanism and the feedback interface based on the gate mechanism are used to guide the model to extract useful patterns and integrate these patterns into the model to enhance spatiotemporal representation respectively. The memory module is used to learn specific knowledge-based patterns, and this is achieved by constraining the learning process with unsupervised loss functions formulated inspired by exogenous knowledge. We construct three kinds of memory modules driven by different exogenous knowledge: the long-term trend memory to learn periodic patterns, the hierarchical effect memory to capture coarse-grained region patterns, and the representative pattern memory to extract representative patterns. Experiments combined with multiple existing models demonstrate the effectiveness of the memory system.