Adaptive Heterogeneous Graph Representation Learning Using kNN-Augmented Graph Mamba Networks (Ka-Gmn)

Abstract

Graph representation learning for heterogeneous networks presents challenges in structural preservation and computational tractability. We present KA-GMN (KNN-Augmented Graph Mamba Networks), integrating k-nearest neighbor selection with state space models for graph representation learning. The architecture implements: (1) KNN-based state transitions for type-specific node representation, (2) compatibility functions for structural graph adaptation, and (3) type-aware feature transformations to prevent representation degradation. KA-GMN processes multi-typed relationships through selective message passing and state space modeling, maintaining graph structure through learned neighborhood functions. The theoretical framework establishes a foundation for heterogeneous graph representation through the synthesis of KNN-based topology and state space dynamics.