Abstract

Neural differential equation models have garnered significant attention inrecent years for their effectiveness in machine learning applications.Amongthese, fractional differential equations (FDEs) have emerged as a promisingtool due to their ability to capture memory-dependent dynamics, which are oftenchallenging to model with traditional integer-order approaches.While existingmodels have primarily focused on constant-order fractional derivatives,variable-order fractional operators offer a more flexible and expressiveframework for modeling complex memory patterns. In this work, we introduce theNeural Variable-Order Fractional Differential Equation network (NvoFDE), anovel neural network framework that integrates variable-order fractionalderivatives with learnable neural networks.Our framework allows for themodeling of adaptive derivative orders dependent on hidden features, capturingmore complex feature-updating dynamics and providing enhanced flexibility. Weconduct extensive experiments across multiple graph datasets to validate theeffectiveness of our approach.Our results demonstrate that NvoFDE outperformstraditional constant-order fractional and integer models across a range oftasks, showcasing its superior adaptability and performance.