Differentiable Logic Rule Induction Algorithms For Interpretable Time Series Forecasting

Abstract

Time series forecasting is one of the most important tools for use cases like finance, energy management, weather prediction, and healthcare. Although recent deep learning models exhibit impressive predictive performances, their inner mechanisms work in black boxes without giving any insight into their logic. To overcome this problem, the study introduces DLRIA (Differentiable Logic Rule Induction Algorithms). It is an innovative method to perform time series forecasting by combining logic rule induction and neural networks. DLRIA uses differentiable fuzzy logic operators and symbolic reasoning to discover human-interpretable logical rules from time series data. In this architecture, there are three main stages, including feature extraction using temporal convolutional networks, logic rule induction using differentiable decision trees and fuzzy membership functions, and interpretability generation using rule extraction and explanation synthesis. The proposed method was experimented on standard datasets, including electricity load, stock prices, traffic flow, and air quality. Experiments have shown that DLRIA can generate comparable forecasting results with only 5-12% lower MAPEs compared to pure neural methods while achieving more than 95% rule interpretability. The proposed system provides concrete rules like “IF temperature rises by 2° Celsius AND humidity is high, THEN electricity consumption rises by 15% ± 3% tomorrow,” which enables domain experts to verify predictions based on known facts. In this way, the DLRIA framework enables time series forecasting to be both accurate and interpretable, making it suitable for implementation in regulated fields like smart grids, health care, and finance.