Tensor-Decomposed Sparse Attention Mechanisms for Scalable Heterogeneous Graph Neural Network Inference on Edge-Constrained Architectures

Abstract

Heterogeneous graph neural networks (HGNNs) have demonstrated considerable efficacy in modeling complex relational structures; however, their deployment on edge-constrained hardware remains critically hindered by prohibitive memory footprints and quadratic attention complexity. This work introduces a tensor-decomposed sparse attention framework (TDSAF) that leverages Tucker decomposition and structured sparsity masks to reduce per-layer parameter density by up to 73% without measurable degradation in downstream node classification accuracy. Benchmark evaluations conducted on OGB-MAG, IMDB, and DBLP heterogeneous graph datasets confirm that TDSAF achieves a 4.1× inference speedup relative to state-of-the-art HGT and HAN baselines under identical floating-point operation budgets. The proposed meta-path-aware pruning scheduler further enables dynamic sparsity adaptation across heterogeneous relation types, advancing the practical deployability of graph-based reasoning on resource-constrained embedded systems.