Interpretability in Deep Time Series Models Demands Semantic Alignment

TL;DR

The paper argues that deep time series models suffer from semantic opacity, where internal representations do not align with human-domain concepts. It formalizes semantic alignment (SA) for temporal data, requiring encoder outputs to match instantaneous concepts and propagation to preserve dynamic concepts over time, while also constraining mechanisms to human-defined relations. A blueprint is proposed to build semantically aligned deep time series models by extending concept-based models to temporally evolving concepts, with a training objective that enforces both spatial and temporal SA and allows residual pathways for expressivity. The work discusses the impact of SA on actionability, verifiability, and robustness, and outlines opportunities for test-time interventions, concept-conditioned generation, and neurosymbolic integration, while addressing counterarguments and practical concerns like annotation cost.

Abstract

Deep time series models continue to improve predictive performance, yet their deployment remains limited by their black-box nature. In response, existing interpretability approaches in the field keep focusing on explaining the internal model computations, without addressing whether they align or not with how a human would reason about the studied phenomenon. Instead, we state interpretability in deep time series models should pursue semantic alignment: predictions should be expressed in terms of variables that are meaningful to the end user, mediated by spatial and temporal mechanisms that admit user-dependent constraints. In this paper, we formalize this requirement and require that, once established, semantic alignment must be preserved under temporal evolution: a constraint with no analog in static settings. Provided with this definition, we outline a blueprint for semantically aligned deep time series models, identify properties that support trust, and discuss implications for model design.

Figures (3)

• Figure 1: Template architecture for time series tasks, characterizing the class of models considered in this work. Observations within a window are first mapped by an Encoder into latent representations, which are transformed by a task-specific Propagation module and subsequently mapped by a Decoder to the target output. Modules operating on latent representations are highlighted in gray.
• Figure 2: Computational graph showing the inference pathway for a model from the class described in Sec. \ref{['sec:DeepTSmodels']}. The symbol $\sim$ indicates SA of representations with concepts.
• Figure 3: Examples of possible instantiations of the blueprint for different time series tasks. Es. (a) Concept-based forecasting: some of the input variables (green and orange) are chosen as the forecasting target, while all available input information is used to encode new dynamic concepts that contribute to the forecasting. Es. (b) Concept-based window classification: instantaneous concepts are extracted from different input subsequences and then combined to solve a downstream task. Es. (c) Concept-based sequence generation: the first input subsequence is used to encode one instantaneous concept and one dynamic concept; later, the second subsequence is used to drive the evolution of the dynamic concept. Finally, all available interpretable information is used to generate a new time series.

Theorems & Definitions (2)

• definition 1: Semantic alignment of concepts
• definition 2: Semantic alignment of mechanisms