HGF-Mixer: A Heterogeneous Graph Fusion Architecture for Spatio-temporal Forecasting

Accurate traffic forecasting is a cornerstone of Intelligent Transportation Systems (ITS). The proliferation of Unmanned Aerial Vehicles (UAVs) provides high-resolution, multi-dimensional traffic data, enabling fine-grained prediction but also introducing significant challenges. State-of-the-art models are often confounded by dual heterogeneity issues: (1) feature heterogeneity, where diverse metrics like vehicle counts and speeds exhibit distinct dynamics, and (2) spatial heterogeneity, where multifaceted correlations between road segments cannot be captured by a single graph. To address these challenges, we introduce the Heterogeneous Graph Fusion Architecture for Spatio-temporal Forecasting (HGF-Mixer), a novel network that explicitly models distinct relation types. HGF-Mixer disentangles complex spatial dependencies by dynamically constructing and integrating three graphs for each input sample: a semantic graph, a pattern-based graph, and an influence graph. At its core, a relation-aware graph convolution module learns dedicated representations for each graph type and employs an attention mechanism to fuse them adaptively, enabling a nuanced understanding of traffic evolution. Extensive experiments on a challenging real-world UAV dataset demonstrate that HGF-Mixer significantly outperforms leading baseline models across multiple forecasting metrics, validating its superiority in handling highfrequency, multi-dimensional traffic data.