Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series

Sachith Abeywickrama, Emadeldeen Eldele, Min Wu, Xiaoli Li, Chau YuenarXiv 2026 · arXiv

TL;DR

ConceptTime is a novel framework for interpretable time-series classification that uses "predictive surprise" from a self-supervised forecaster to simultaneously segment signals and define meaningful concepts. It achieves state-of-the-art accuracy while providing built-in reliability scores and human-inspectable decision paths, outperforming existing interpretable models across 17 diverse benchmarks.

What Problem Does This Paper Address?

In risk-sensitive domains like healthcare or industrial monitoring, AI models face a difficult "trifecta" of requirements: they must be highly accurate, human-interpretable, and capable of "calibrated abstention" (knowing when to say "I don't know").

Existing methods struggle with a structural "chicken-and-egg" problem unique to sequential data: you cannot define a meaningful temporal pattern (a "concept") without knowing where its boundaries are, but you cannot place those boundaries accurately without knowing what patterns to expect. Most current approaches sidestep this by using fixed-length windows or expensive, manually annotated vocabularies, which often leads to a sacrifice in at least one of the three core requirements.

Methods

The authors introduce ConceptTime, which closes this loop using a single self-supervised signal.

  1. Self-Supervised Pretraining: They train an Autoregressive Gaussian Density Model (ARGDM)—a 4-layer causal Transformer—on unlabeled raw signals to predict the probability distribution of the next observation at every timestep.
  2. Predictive Surprise: The mismatch between the model's prediction and the actual realized signal is called "predictive surprise" (measured as negative log-likelihood). High surprise naturally indicates a regime change, which the model uses to draw segment boundaries via dynamic programming.
  3. Grounded Signatures: Instead of just looking at raw shapes, each segment is characterized by its "predictive geometry"—a signature vector combining predictive moments (what was expected), residual morphology (how the error looked), and the surprise trajectory (how the error unfolded).
  4. Concept Bottleneck: These signatures are clustered into a small vocabulary of prototypes (concepts) using a Gaussian Mixture Model (GMM). A lightweight classifier then makes decisions based on the sequence of these concepts.
  5. Fidelity Score: The distance from a new segment to its nearest learned prototype yields a built-in "fidelity score." This allows the system to abstain from a prediction if the input signal is fundamentally different from anything in its learned vocabulary.
The ConceptTime pipeline: Stage 1 learns concepts from unlabeled data; Stage 2 applies them to downstream classification with built-in fidelity scores.

Datasets

The model was evaluated across 17 public benchmarks spanning four domains:

  • Wearable Sensing: UCI HAR (Human Activity Recognition).
  • Biomedical: UCI Epilepsy (seizure detection) and Sleep-EDF (sleep staging).
  • General Time Series: 10 datasets from the UEA multivariate archive (ranging from handwriting to heartbeat monitoring).
  • Industrial Telemetry: Fault Diagnosis (FD) datasets used specifically to test cross-domain transferability.

Key Insights and Ideas

  • Behavior Over Shape: A key "aha" moment is that grounding concepts in behavior relative to expectation is more discriminative than grounding them in raw shape. As shown in Figure 1, two identical-looking flat signals might represent different concepts: one might be an expected "calm" period, while the other is a surprising "lull" following a high-intensity spike.
  • The Dual-Use Signal: The paper demonstrates that the same self-supervised signal used for segmentation (surprise) is also ideal for feature extraction. This eliminates the need for human annotations or external language models to define concepts.
  • Built-in Reliability: Unlike standard "black-box" models that might give a high-confidence wrong answer, ConceptTime’s fidelity score asks a more fundamental question: "Is this input even describable in my concept space?" This makes its abstention mechanism much more robust for OOD (Out-of-Distribution) detection.
Figure 1: ConceptTime targets accuracy, interpretability, and calibrated abstention through a single self-supervised primitive. Left, existing methods emphasize different goals. Right, 4 segments (A-D) and their preceding context. Raw-shape space conflates A and B (both end calm) and splits C from D. Predictive-geometry space, spanned by predictive variance and NLL, collapses A, C, and D as expectation is satisfied and isolates B as a surprise. Concepts grounded in deviation from prediction, therefore, distinguish dynamical regimes that shape-based vocabularies cannot.

Key Results

  • Accuracy & Efficiency: ConceptTime is competitive with state-of-the-art black-box models like ROCKET and TimesNet. Remarkably, it retains near-98% accuracy on the Epilepsy dataset using only 1% of the available labels, demonstrating extreme sample efficiency.
  • Interpretable Superiority: It beats every interpretable baseline tested, often by massive margins (e.g., a 45-point accuracy lead over raw-signal morphology on the UCI-HAR dataset).
  • OOD Detection: The fidelity score achieves an AUROC of up to 0.99 for OOD detection, significantly outperforming standard softmax confidence scores which often remain near chance levels on perturbed data.

Limitations and Future Work

The authors acknowledge a few areas for growth:

  • Fixed Vocabulary: The number of concepts ($M$) must be chosen in advance. Future iterations could use hierarchical vocabularies to capture patterns at multiple levels of detail.
  • Human Interpretation: While the concepts are discovered automatically, their semantic meaning (e.g., "this concept represents a pre-seizure spike") still requires post-hoc human interpretation.
  • Static Prototypes: The current model uses a frozen vocabulary. Online updates could help the model adapt as signal patterns drift over time in long-term monitoring scenarios.

Summary updated

Abstract

Time-series classifiers deployed in risk-sensitive domains must be (1) accurate, (2) interpretable, and (3) able to abstain when evidence is weak. Existing methods do not deliver all 3 at once, and the obstacle is structural, rather than logistical. Meaningful temporal patterns cannot be defined without knowing segment boundaries, yet meaningful boundaries cannot be placed without knowing which patterns to expect. Existing approaches break this circularity by using fixed windows or expert-annotated vocabularies, thereby sacrificing at least one of the 3 properties. We introduce ConceptTime, which closes the loop through a single self-supervised signal. A frozen probabilistic forecaster predicts a distribution over the next observation at every timestep of the sequence. We call the mismatch between its prediction and the realized signal the predictive surprise. This signal both locates segment boundaries and characterizes each resulting segment through the forecaster. Segments are thus grounded by how they behaved relative to expectation, not by what they look like. These summaries cluster into a vocabulary of dynamical regimes that we call concepts. A lightweight head classifies from the concept sequence, and the distance to the nearest concept yields reliability by design. Across 17 public benchmarks, ConceptTime is competitive with state-of-the-art black-box accuracy, beats every interpretable baseline, and retains near-98% accuracy on Epilepsy with 1% of labels. It further remains competitive with softmax, prototype-distance, and SHAP/TimeX/TimeX++ on OOD detection and deletion faithfulness, without a single annotation, language model call, or domain expert. Code is available at: https://github.com/Sachithx/ConceptTime.git.
BibTeX
@misc{abeywickrama2026predictive,
  author = {Sachith Abeywickrama and Emadeldeen Eldele and Min Wu and Xiaoli Li and Chau Yuen},
  title = {{Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series}},
  year = {2026},
  eprint = {2611.12955},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.12955},
}
Figure