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

Predictive Surprise as Self-Grounding Concept Bottleneck for Interpretable Time Series. By Sachith Abeywickrama and four others, from Nanyang Technological University, A-STAR, and two other institutions. Published on the seventh of November.

A time-series classifier used in a risk-sensitive domain has to do three things: be accurate, be interpretable, and abstain when the evidence is weak. Existing methods do not manage all three together, and the authors argue the obstacle is structural rather than logistical. It is a circle. Meaningful temporal patterns cannot be defined without knowing where segments begin and end, and meaningful segment boundaries cannot be placed without knowing which patterns to expect. Current approaches break the circle with fixed windows or with vocabularies annotated by experts, and in doing so they give up at least one of the three properties. A fixed window places boundaries without regard to the patterns. An expert vocabulary decides the patterns in advance, at the cost of the annotation it requires. Either way, one side of the circle is simply assumed.

ConceptTime closes the loop with a single self-supervised signal. A frozen probabilistic forecaster predicts a distribution over the next observation at every timestep. The mismatch between that prediction and what the signal then actually does is named the predictive surprise. That one quantity does both jobs the circle needed: it locates segment boundaries, and, through the forecaster, it characterizes each resulting segment. A segment is grounded in how it behaved relative to expectation, not in what it looks like. That is the sense of self-grounding in the title. The forecaster is frozen and the signal is self-supervised, so neither the boundaries nor the descriptions need anyone to label them.

The segment summaries are clustered into a vocabulary of dynamical regimes, which the authors call concepts. A lightweight head then classifies from the sequence of concepts. Abstention falls out of the same structure: the distance from a segment to its nearest concept gives a measure of reliability by design, with no separate machinery bolted on for it.

Across seventeen public benchmarks, ConceptTime is competitive with state-of-the-art black-box accuracy and beats every interpretable baseline. On Epilepsy, with one percent of the labels, it keeps accuracy near ninety-eight percent.

It also stays competitive with softmax, prototype-distance, and SHAP, TimeX and TimeX++ on two further tests: out-of-distribution detection and deletion faithfulness. None of this uses a single annotation, a language model call, or a domain expert.

Instead of choosing which of the three properties to sacrifice, ConceptTime derives all three from surprise. Boundaries come from where the forecaster was wrong, concepts from how it was wrong, and reliability from how far a segment sits from the concepts it has learned. The vocabulary that experts used to supply is grown from the data's own departures from expectation.