Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Inverse Cross-Spectral Neural Networks (iCSNNs) are introduced to address the limitations of existing CoVariance Neural Networks in capturing joint temporal and cross-variable dependencies in stationary multivariate time series. By leveraging the inverse cross-spectral density (iCSD) matrices as graph shift operators, iCSNNs encode frequency-specific conditional relationships, exploiting spectral smoothness through band grouping for compact parametrization. Empirical results on synthetic data demonstrate iCSNN's superior performance over established baseline methods.
Frequency-specific conditional relationships in multivariate time series are unlocked by using inverse cross-spectral density matrices as graph shift operators.
CoVariance Neural Networks and their extensions have emerged as effective tools for processing multivariate data, deriving graph shift operators directly from second-order statistics. These architectures, however, are designed for independent and identically distributed observations and do not fully capture the joint structure of temporal and cross-variable dependencies in multivariate time series. In this work, we introduce Inverse Cross-Spectral Neural Networks (iCSNNs), a class of graph neural networks for stationary multivariate time series whose shift operators are the inverse cross-spectral density (iCSD) matrices. These operators encode frequency-specific conditional relationships among variables, exploiting the decomposition provided by the spectral representation theorem. Leveraging spectral smoothness, frequencies are grouped into bands sharing a single iCSD operator, yielding a compact parametrisation that retains the frequency-dependent structure of the process. We further propose a joint learning procedure to estimate both the Fourier-domain dependence structure and the iCSNN parameters, adapting the iCSD operators to the downstream task. When tested on synthetic data, iCSNN outperforms baselines from different methodological families.