Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series
Time-series classifiers are often asked to do three difficult things at once: make accurate predictions, explain which parts of a signal mattered, and recognize when they should abstain. These demands are especially important in healthcare, wearable sensing, industrial monitoring, and other settings where a confident mistake can be costly. The authors of Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series argue that existing methods struggle because time-series interpretation has a structural chicken-and-egg problem.
A useful concept depends on knowing where a meaningful segment begins and ends. Yet meaningful boundaries are difficult to choose without knowing which patterns matter. Fixed windows solve the first problem but may cut across important events. Expert-defined concepts provide semantics but are expensive and incomplete. Shape-based prototypes can be interpretable while ignoring the signal’s preceding context. The paper introduces ConceptTime as an attempt to resolve this circularity with one self-supervised signal: predictive surprise.
Using prediction errors to find temporal regimes
ConceptTime first trains an autoregressive Gaussian density model on unlabeled time-series data. At every timestep, the model predicts a distribution for the next observation, including a mean and an uncertainty estimate. After training, the model is frozen. Predictive surprise is the negative log-likelihood of the observation that actually arrives under that predicted distribution.
This distinction matters. A surprising observation is not merely one with high predicted variance. It is an observation that differs from what the model expected, particularly when the model was confident. The authors use this per-timestep surprise signal to place segment boundaries. A length-constrained dynamic-programming procedure selects a target number of non-overlapping patches, favoring boundaries around concentrated violations of expectation while preventing degenerate segments.
The same model then describes each patch. ConceptTime combines predictive moments, statistics of standardized residuals, and the trajectory of surprise within the segment. These features capture what the model expected, how the residual behaved, and when the unexpected behavior occurred. Two waveforms that look similar in isolation can therefore receive different representations if one is expected in context and the other is not.
From signatures to inspectable concepts
The patch signatures are clustered with a Gaussian mixture model to produce a vocabulary of concepts. The concepts are not human-labeled medical or engineering terms during training. Instead, each is a prototype in predictive-geometry space, with nearby training patches serving as examples that people can inspect afterward.
A lightweight classifier receives the sequence of soft concept assignments rather than the raw signal. This makes the decision route more explicit: signal, adaptive segments, predictive signatures, concept assignments, and finally the class prediction. The paper emphasizes that the signal-to-concept path is deterministic and label-free; labeled supervision is used only for the downstream classifier.
Prototype distance also supplies a built-in fidelity score. High fidelity means that the patches lie close to at least one learned prototype. Low fidelity means that part of the input falls in a region poorly covered by the vocabulary. This is different from softmax confidence, which asks whether the classifier favors one class. Fidelity instead asks whether the input is represented by concepts the system has actually learned.
What the experiments showed
The authors evaluated ConceptTime across 17 public benchmarks covering wearable sensing, EEG and sleep data, multivariate classification, and bearing-fault diagnosis. They report that ConceptTime was the strongest interpretable method across the main classification comparisons and remained competitive with black-box systems, although the cost of enforcing a concept bottleneck was visible on some fully supervised tasks.
The clearest result came from the Epilepsy benchmark. ConceptTime-Dynamic reached 98.39% accuracy with full supervision and 97.91% using 1% of the labels in the reported table. On UCI-HAR and Sleep-EDF, black-box baselines performed better in several settings, so the method should not be described as universally more accurate. The authors’ ablations indicate that the benefit comes from predictive geometry rather than adaptive segmentation alone: replacing predictive signatures with raw-signal morphology substantially reduced performance, including a reported 45-point gap on UCI-HAR in one matched comparison.
The approach also showed useful transfer behavior on four fault-diagnosis operating conditions. With 5% of target labels and a target-trained density model while retaining the source vocabulary, mean accuracy across 12 transfer directions reached 94.37%, compared with 91.09% when using a source density model and 90.76% for the cited TSLANet comparison. These are specific benchmark results, not evidence that the vocabulary will transfer unchanged across arbitrary industries or sensors.
For selective prediction, fidelity detected synthetic perturbations on human-activity data strongly in the reported experiments, with AUROC ranging from 0.864 to 0.992 as noise severity increased. On real operating-condition shifts in fault diagnosis, however, the mean AUROC was 0.666 and the false-positive rate at 95% recall was 0.719. That contrast is important: the score can be useful, but domain shift remains difficult.
Practical implications and limitations
ConceptTime offers a practical design pattern for teams with abundant unlabeled signals but limited annotations. A forecaster can be trained once, then reused to create adaptive segments, discover a compact prototype vocabulary, support few-shot classification, and flag poorly represented inputs for review. In an industrial setting, this could help organize recurring operating regimes. In biomedical monitoring, it could provide segment-level evidence for inspection rather than presenting only a class probability.
Still, a prototype is not automatically a meaningful human concept. Its interpretation is post hoc, the vocabulary size must be chosen in advance, and a high fidelity score does not guarantee a correct prediction. The authors explicitly warn that prototypes may capture sensor artifacts as easily as clinically or mechanically meaningful events. Thresholds for abstention must therefore be validated on the intended population, sensors, and operating conditions, with human oversight for consequential decisions.
The paper’s central contribution is less a promise of universally superior accuracy than a way to connect segmentation, representation, interpretation, and abstention. Predictive surprise provides a shared reference point: it identifies where behavior changes and describes how that behavior differs from expectation. Whether that foundation produces trustworthy concepts in a new domain will depend on the quality of the forecaster, the coverage of the unlabeled data, and careful evaluation beyond benchmark accuracy.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.