NFCL: Simply interpretable neural networks for a short-term multivariate forecasting

TL;DR

NFCL tackles the challenge of interpretable short-term multivariate time-series forecasting by proposing a simple additive layer that assigns independent inputs-to-outputs contributions. It presents three expansions—Simple (NFCL-V), Complex (NFCL-C), and Decomposed (NFCL-D)—that progressively introduce nonlinearity, per-input mappings, and multiscale decomposition, with NFCL-C showing the best overall performance. Across 15 datasets and nine baselines, NFCL achieves strong metrics, particularly SMAPE and R^2, while providing explicit contribution maps that explain forecasts. The work demonstrates that interpretability can coexist with competitive accuracy, albeit with considerations around data size and computational complexity, and points to practical deployment opportunities such as web-based tools for non-experts.

Abstract

Multivariate time-series forecasting (MTSF) stands as a compelling field within the machine learning community. Diverse neural network based methodologies deployed in MTSF applications have demonstrated commendable efficacy. Despite the advancements in model performance, comprehending the rationale behind the model's behavior remains an enigma. Our proposed model, the Neural ForeCasting Layer (NFCL), employs a straightforward amalgamation of neural networks. This uncomplicated integration ensures that each neural network contributes inputs and predictions independently, devoid of interference from other inputs. Consequently, our model facilitates a transparent explication of forecast results. This paper introduces NFCL along with its diverse extensions. Empirical findings underscore NFCL's superior performance compared to nine benchmark models across 15 available open datasets. Notably, NFCL not only surpasses competitors but also provides elucidation for its predictions. In addition, Rigorous experimentation involving diverse model structures bolsters the justification of NFCL's unique configuration.

Figures (9)

• Figure 1: Forecasting the fourth and seventh next steps' HUFL based on the historical multivariate time-series. Below, contribution maps illustrate the reasons why the model predicts the target to be closer to $\hat{Y}^{\mathrm{HUFL}}=10$. The pattern of the contribution map predicted by the model changes, similar to moving three steps forward in the fourth future step while the target time point is moved to three step forward.
• Figure 2: The simple expansion for forecasting. The black arrow denotes a fully-connected operation between each input sequence in X and each target sequence in Y.
• Figure 3: The complex expansion for forecasting.
• Figure 4: The decomposition for forecasting.
• Figure 5: The NFCLs overall pipeline for forecasting.