SpikF-GO brings graph-based modeling to spiking neural networks for time series

Spiking Neural Networks (SNNs) are known for energy efficiency, but their application to time series forecasting (TSF) has been limited by a critical flaw: existing SNN methods treat each variable independently, ignoring the cross-variable correlations that are essential for multivariate forecasting. Now, a new paper from Jafar Bakhshaliyev and Niels Landwehr, accepted at ECML PKDD 2026, introduces SpikF-GO (Spiking Fourier Graph Operators) to close that gap. SpikF-GO models every scalar observation as a node in a hypervariate graph, enabling spike-driven spectral processing that captures dependencies across all variables.

The architecture introduces two novel components: a Hard Concrete frequency gate that learns to select only the most informative frequencies (sparse selection), and a Complex LIF (Leaky Integrate-and-Fire) gate that applies independent spiking neurons to the real and imaginary parts of Fourier components, preserving binary, event-driven computation in the frequency domain. A variant also uses Central Pattern Generator (CPG)-based positional encodings for better long-range temporal modeling. Tested on eight benchmark datasets under a unified protocol, SpikF-GO achieved the best average rank among all SNN methods and outperformed the equivalent ANN model (FourierGNN) while consuming less energy. Notably, it maintains competitive accuracy even with much smaller embedding dimensions, further reducing energy costs. This work is among the first to bring graph-based multivariate modeling into the spiking domain for TSF and provides the first unified comparison of SNN forecasting architectures.