New SNN Method Improves Multivariate Time Series Forecasting with Graph Operators.

A new research paper introduces SpikF-GO, an innovative Spiking Neural Network (SNN) architecture designed for multivariate time series forecasting. Unlike previous SNN methods that process variables independently, SpikF-GO incorporates a hypervariate graph formulation, treating each scalar observation as a graph node, to explicitly model complex inter-variable dependencies. The core of SpikF-GO involves spike-driven spectral processing, utilizing a Hard Concrete frequency gate for sparse frequency selection and a Complex LIF gate for handling real and imaginary Fourier components with independent spiking neurons. This design maintains binary, event-driven computation within the spectral domain. Evaluations across eight benchmarks show that SpikF-GO achieves the best average rank among SNN methods and surpasses its ANN counterpart, FourierGNN, while significantly reducing energy consumption, even with smaller embedding dimensions. This work is notable for bringing graph-based multivariate modeling to the spiking domain for time series forecasting.