Curvature-Informed Attention Mechanism for Long Short-Term Memory Networks

Abstract

Time series forecasting is a crucial task across diverse domains, and recent research focuses on refining model architectures for enhanced predictive capabilities. In this paper, we introduce a novel approach by integrating curvature measures into an attention mechanism alongside Long Short-Term Memory (LSTM) networks. The objective is to improve the interpretability and overall performance of time series forecasting models. The proposed Curvature-Informed Attention Mechanism (CIAM) enhances learning by personalizing the weight attribution within the attention mechanism. Through comprehensive experimental evaluations on real-world datasets, we demonstrate the efficacy of our approach, showcasing competitive forecasting accuracy compared to traditional LSTM models.