TimeSliver : Symbolic-Linear Decomposition for Explainable Time Series Classification

TL;DR

A novel explainability-driven deep learning framework, TimeSliver, which jointly utilizes raw time-series data and its symbolic abstraction to construct a representation that maintains the original temporal structure, positioning it as a strong and explainable framework for general time-series classification.

Abstract

Identifying the extent to which every temporal segment influences a model's predictions is essential for explaining model decisions and increasing transparency. While post-hoc explainable methods based on gradients and feature-based attributions have been popular, they suffer from reference state sensitivity and struggle to generalize across time-series datasets, as they treat time points independently and ignore sequential dependencies. Another perspective on explainable time-series classification is through interpretable components of the model, for instance, leveraging self-attention mechanisms to estimate temporal attribution; however, recent findings indicate that these attention weights often fail to provide faithful measures of temporal importance. In this work, we advance this perspective and present a novel explainability-driven deep learning framework, TimeSliver, which jointly utilizes raw time-series data and its symbolic abstraction to construct a representation that maintains the original temporal structure. Each element in this representation linearly encodes the contribution of each temporal segment to the final prediction, allowing us to assign a meaningful importance score to every time point. For time-series classification, TimeSliver outperforms other temporal attribution methods by 11% on 7 distinct synthetic and real-world multivariate time-series datasets. TimeSliver also achieves predictive performance within 2% of state-of-the-art baselines across 26 UEA benchmark datasets, positioning it as a strong and explainable framework for general time-series classification.

Figures (10)

• Figure 1: Overview of TimeSliver: (Module I) temporal segment extraction and latent representation learning ($\mathbf{g}(\mathbf{x}_i; \theta_{\mathrm{q}}$)); (Module II) symbolic composition of temporal segments; and (Module III) global linear interaction between latent and symbolic representations to generate $\boldsymbol{P}$, a representation of $\mathbf{x}_i$ preserving temporal structure. $\boldsymbol{P}$ is then passed through a linear layer ($\mathbf{h}(\mathbf{x}_i; \theta_{\mathrm{c}}$)) to predict $y_i$ and used to compute temporal attribution using $f_{att}$. The right column compares ground truth attribution scores with baseline methods and TimeSliver, where darker regions indicate positive influence.
• Figure 1: Comparison of $\text{mean}_{\pm \text{std}}$ AUPRC on synthetic datasets. Bold: best, underlined: second-best. NA: not applicable
• Figure 2: (a) shows a raw time series input, (b) is the symbolic composition matrix, $\boldsymbol{Z}$ and (c) shows some sample rows of $\boldsymbol{Z}$ which serve as the Bag-of-Stencils to modulate $\boldsymbol{P}$.
• Figure 3: Impact of using raw $\boldsymbol{X}$ instead of $\boldsymbol{Z}$ on (a) explainability (AUPRC) and (b) predictability (Accuracy).
• Figure 4: Positive attribution study. Accuracy curves $e(u)$ plotted against the unmasking percentage $u$% for EEG dataset.

Theorems & Definitions (2)

• Definition 2.1: Temporal Attribution-Based Explainability
• Definition 2.2: Temporal Segment