Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Self-explainable Model of Long Time Series by Extracting Informative Structured Causal Patterns

Ziqian Wang, Yuxiao Cheng, Jinli Suo

TL;DR

<3-5 sentence high-level summary> EXCAP tackles the challenge of making long time-series models explainable by jointly learning temporally coherent segment-level explanations and causally grounded predictions. It introduces a three-component architecture—attention-based segmenter, encoder, and interpretable causal decoder guided by a pre-trained graph—augmented with latent-space aggregation losses and a staged optimization schedule. The approach provides theoretical guarantees on temporal continuity, faithfulness, and efficiency, and demonstrates superior interpretability alongside competitive predictive performance on classification and forecasting benchmarks. The framework scales linearly with sequence length and holds promise for high-stakes domains requiring trustworthy and actionable explanations.

Abstract

Explainability is essential for neural networks that model long time series, yet most existing explainable AI methods only produce point-wise importance scores and fail to capture temporal structures such as trends, cycles, and regime changes. This limitation weakens human interpretability and trust in long-horizon models. To address these issues, we identify four key requirements for interpretable time-series modeling: temporal continuity, pattern-centric explanation, causal disentanglement, and faithfulness to the model's inference process. We propose EXCAP, a unified framework that satisfies all four requirements. EXCAP combines an attention-based segmenter that extracts coherent temporal patterns, a causally structured decoder guided by a pre-trained causal graph, and a latent aggregation mechanism that enforces representation stability. Our theoretical analysis shows that EXCAP provides smooth and stable explanations over time and is robust to perturbations in causal masks. Extensive experiments on classification and forecasting benchmarks demonstrate that EXCAP achieves strong predictive accuracy while generating coherent and causally grounded explanations. These results show that EXCAP offers a principled and scalable approach to interpretable modeling of long time series with relevance to high-stakes domains such as healthcare and finance.

A Self-explainable Model of Long Time Series by Extracting Informative Structured Causal Patterns

TL;DR

<3-5 sentence high-level summary> EXCAP tackles the challenge of making long time-series models explainable by jointly learning temporally coherent segment-level explanations and causally grounded predictions. It introduces a three-component architecture—attention-based segmenter, encoder, and interpretable causal decoder guided by a pre-trained graph—augmented with latent-space aggregation losses and a staged optimization schedule. The approach provides theoretical guarantees on temporal continuity, faithfulness, and efficiency, and demonstrates superior interpretability alongside competitive predictive performance on classification and forecasting benchmarks. The framework scales linearly with sequence length and holds promise for high-stakes domains requiring trustworthy and actionable explanations.

Abstract

Explainability is essential for neural networks that model long time series, yet most existing explainable AI methods only produce point-wise importance scores and fail to capture temporal structures such as trends, cycles, and regime changes. This limitation weakens human interpretability and trust in long-horizon models. To address these issues, we identify four key requirements for interpretable time-series modeling: temporal continuity, pattern-centric explanation, causal disentanglement, and faithfulness to the model's inference process. We propose EXCAP, a unified framework that satisfies all four requirements. EXCAP combines an attention-based segmenter that extracts coherent temporal patterns, a causally structured decoder guided by a pre-trained causal graph, and a latent aggregation mechanism that enforces representation stability. Our theoretical analysis shows that EXCAP provides smooth and stable explanations over time and is robust to perturbations in causal masks. Extensive experiments on classification and forecasting benchmarks demonstrate that EXCAP achieves strong predictive accuracy while generating coherent and causally grounded explanations. These results show that EXCAP offers a principled and scalable approach to interpretable modeling of long time series with relevance to high-stakes domains such as healthcare and finance.

Paper Structure

This paper contains 47 sections, 4 theorems, 22 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Explainable Artificial Intelligence (XAI)
  4. Explainability for Time-Series Models
  5. Causal and Prototype-Based Interpretability
  6. The Proposed EXCAP Framework
  7. Overview of the Proposed Framework
  8. Attention-Based Segmenter
  9. Reference model.
  10. Attention-guided segmentation.
  11. Encoder
  12. Interpretable Decoder
  13. Decoder.
  14. Causal disentanglement.
  15. Segment attribution and contrastive learning.
  16. ...and 32 more sections

Key Result

Proposition 1

Assume that the EXCAP explanation mapping $g(\cdot)$ is composed of a sequence of operators: instance normalization $\mathcal{N}(\cdot)$, attention-based segmentation $\mathcal{A}(\cdot)$, 1D max pooling $\mathcal{P}(\cdot)$, change-point detection $\mathcal{C}(\cdot)$, and a ReLU-based temporal con is $C$-Lipschitz with constant Consequently, for any bounded perturbation $\Delta \mathbf{X}$,

Figures (3)

  • Figure 1: Overview and comparative positioning of EXCAP. (a) The proposed EXCAP framework extracts temporally and causally structured representations. (b) Qualitative comparison of interpretability properties across representative approaches.
  • Figure 2: Architecture of the EXCAP framework. The attention-based segmenter divides the input multivariate time series into temporally coherent segments; the encoder maps each segment into a latent representation capturing both local and global dynamics; the causal decoder integrates these embeddings under a pre-trained causal graph to yield disentangled predictions.
  • Figure 3: Perturbation-based interpretability evaluation. (a) Attribution map visualization on Epilepsy. EXCAP highlights coherent seizure-relevant motifs. (b) Performance degradation under increasing masking ratios on MITECG. Larger AUPRC/AUROC drop $\Rightarrow$ higher attribution faithfulness. (c) Comparison between masking high-attribution vs. low-attribution regions. (d) EXCAP attention on seizure/non-seizure EEG samples, aligned with clinically meaningful spike-wave bursts. (e) t-SNE of EXCAP latent space: high-attention segments cluster by waveform archetypes, while background segments remain diffuse.

Theorems & Definitions (8)

  • Proposition 1: Global Lipschitz Stability of Explanations
  • proof : Sketch of proof
  • Proposition 2: Deletion Faithfulness Lower Bound
  • proof : Sketch of proof
  • Proposition 3: Segment-Level Pattern Preservation
  • proof
  • Proposition 4: Linear-Time Complexity
  • proof