Interpretable deep convolutional model for nonlinear multivariate time series in complex systems.

Area of Science:

• Machine Learning
• Time Series Analysis
• Deep Learning

Background:

• Standard time series forecasting models often act as black boxes, limiting interpretability.
• Understanding the underlying interaction structure in nonlinear multivariate time series is crucial for accurate forecasting.

Purpose of the Study:

• To introduce the Deep Convolutional Interpreter for Time Series (DCIts), a deep learning architecture.
• To provide sample-specific, locally interpretable descriptions of interaction structures in nonlinear multivariate time series.
• To achieve competitive forecasting accuracy while prioritizing intrinsic interpretability.

Main Methods:

• DCIts learns a time- and lag-dependent transition tensor, factorized into a Focuser and a Modeler.
• A sparse masking mechanism in the Focuser selects relevant series and time lags.
• Convolutional filters capture temporal and cross-variable dependencies, mapped through a bottleneck network.

Main Results:

• DCIts recovers stable, signed, lag-resolved interaction patterns on benchmark datasets.
• The model achieves competitive forecasting error compared to interpretable baselines.
• The decomposition provides local lag-adjacency structure and signed source-lag contributions for each forecast.

Conclusions:

• DCIts offers a framework for intrinsically interpretable time series forecasting.
• The model successfully balances forecasting accuracy with the need for understandable results.
• DCIts enables direct inspection of effective connectivity and higher-order contributions in time series data.