Dynamic sequential neighbor processing: A liquid neural network-inspired framework for enhanced graph neural networks

Abstract

Integrating information from multi-order neighborhoods is a fundamental strategy in Graph Neural Networks (GNNs) for capturing higher-order structural patterns and enhancing the expressive power of node representations. However, most existing GNNs treat neighbors from different orders as unordered sets and integrate them using static or parallel strategies, thus overlooking the sequential and evolving nature of neighborhood expansion. To address this limitation, we propose a novel GNN framework, SL, which integrates Serialized Neighbor Features with Liquid Neural Networks (LNNs) to enable order-aware, dynamic adaptation of neighbor influence. By modeling neighbor features as ordered sequences and leveraging LNNs' internal feedback dynamics, SL adapts feature extraction in real time based on local context and propagation history. This design offers fine-grained control over hierarchical dependencies and allows dynamic modulation of contributions from different neighborhood layers. SL is model-agnostic and can be seamlessly integrated with both classical and state-of-the-art GNNs. Extensive experiments across ten benchmark datasets show that SL consistently improves node classification accuracy and significantly alleviates over-smoothing in deep GNNs. These results highlight that order-aware and dynamically regulated propagation represents a powerful, flexible alternative to traditional multi-order aggregation, enhancing the adaptability and expressiveness of GNNs for complex graph learning tasks.