TTF-GNN: Memory-Efficient GNNs via Tensor Train Decomposition and Network Folding

Graph Neural Networks (GNNs) have successfully processed graph-structured data across various applications, including recommendation systems, molecular property prediction, and fraud detection. However, increasing the number of GNN layers to capture long-range dependencies leads to excessive memory consumption and over-smoothing, where node representations become indistinguishable. These challenges hinder efficient deployment, particularly in resource-constrained environments. To address this issue, we propose TTF-GNN, a novel framework that integrates Tensor Train Decomposition (TTD) for memory-efficient compression and network folding to mitigate over-smoothing. TTD reduces the size of GNN weight matrices while preserving expressive power, and network folding restructure deep GNNs into recurrent-like blocks, enabling effective depth reduction without performance degradation. Our method significantly reduces memory usage while maintaining or even improving inference accuracy. Experimental results demonstrate that TTF-GNN improves accuracy-memory footprint trade-offs, with improved accuracy despite aggressive compression, e.g., 40x reduction on the Citeseer dataset. Thanks to its significantly reduced parameters, our approach suits real-time, embedded, and parallelized AI systems targeting low-power and high-performance applications. These findings open new avenues for deploying resource-efficient GNN accelerators and integrating GNN inference within edge devices.