Feature-Fusion Module Extensibility

Background: Developing scalable frameworks to effectively harness the predictive capabilities of univariate time series foundation models (Uni-TSFMs) for complex multivariate forecasting tasks is an ongoing challenge in time series analysis.

Question / Future Work: The paper demonstrates DualWeaver’s success by replacing the feature-fusion module $f(\cdot)$ with CNN-based alternatives, suggesting that the framework is extensible. Future work should explore replacing $f(\cdot)$ with other complex multivariate backbones, such as Graph Neural Networks (GNNs) or Transformer-based models, to see if performance gains can be further improved depending on the characteristics of cross-variable correlations in the dataset.