Title: Semantic Alignment is Essential for Interpretable Deep Time Series Models

Abstract:

Although deep time series models are steadily enhancing their predictive capabilities, their practical adoption is hindered by their opacity as black-box systems. Current interpretability methods within this domain predominantly concentrate on elucidating internal model computations, often neglecting to verify whether these explanations resonate with human reasoning regarding the phenomenon under investigation. We argue that interpretability in deep time series modeling must instead prioritize semantic alignment. This approach dictates that predictions be articulated through variables that hold significance for the end user, facilitated by spatial and temporal mechanisms that accommodate user-specific constraints. This paper formalizes this necessity and posits that, once established, semantic alignment must remain consistent over time—a requirement distinct from static contexts. Building on this definition, we propose a framework for developing semantically aligned deep time series models, highlight characteristics that foster trust, and explore the resulting implications for model architecture.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC