Interpretable Fuzzy Embedded Neural Network for Multivariate Time-Series Forecasting

Interpretability in Deep Learning has become a critical component in applied AI research. When it comes to understanding deep learning in time-series contexts, many approaches emphasize visualization methods and post-hoc techniques. Conversely, the EcFNN approach integrates a deep learning model with a fuzzy logic system. This system generates fuzzy rules to unveil the black-box nature of the decision-making of the embedded neural network. These linguistic fuzzy rules are simpler for humans to understand. However, the EcFNN does not support multivariate time-series problems. In this paper, we develop a method called E-EcFNN that supports multivariate time-series problems. Notably, our experiments indicate that our new method provides interpretability and maintains a competitive level of accuracy compared to other baselines.