Facial-Based Heterogeneous Graph Representation Learning for Autism Spectrum Disorder Detection

Autism spectrum disorder (ASD) is a neurodevelopmental disorder whose early manifestations may include atypical facial behaviors. Prior facial-based ASD studies often focus on localized static cues or short-range temporal variations, while under-exploring the joint modeling of temporal dynamics and non-local semantic consistency across facial states. In this work, we propose a graph representation learning framework consisting of a Graph Construction module (GC) and a heterogeneous graph neural network, HGMoE. The GC module builds a temporal–semantic heterogeneous graph from facial videos, where video frames are treated as nodes and heterogeneous edges encode both local temporal continuity and semantic similarity across frames. To better model heterogeneous relations and capture diverse graph signals, HGMoE integrates a Mixture-of-Experts (MoE) architecture with relation-aware edge attention. Specifically, the MoE dynamically routes messages to specialized experts, while the edge attention mechanism explicitly leverages edge types and initial semantic weights to modulate message passing across relations. Experiments on our recruited dataset achieve a classification accuracy of 97.4%, demonstrating the potential of the proposed framework for supporting facial-based ASD screening.