Temporal Prototype Attribution

The model learns a set of prototypical time-series subsequences during training that capture the fundamental patterns of each class or forecast regime. These prototypes are not raw training samples but learned, optimized representations that reside in the model's latent space. The objective is to encode the data manifold into a small, interpretable set of exemplars that are maximally representative of the underlying dynamics.

A prediction is explained by quantifying the similarity between the input sequence and each learned prototype. The model computes a distance metric—often Euclidean or a learned similarity function—in the latent representation space. The final output is attributed to the prototypes with the highest similarity scores, providing a case-based reasoning trace. This directly answers: 'This forecast was made because the recent data looks like these known patterns.'

Unlike methods that highlight individual time steps, this approach identifies entire subsequences as the unit of explanation. The model decomposes an input time series into parts and matches each part to the closest prototype. This provides a structural explanation, showing which segments of the input are driving the decision and which canonical pattern they resemble. It naturally handles variable-length temporal dependencies.

Explanations are generated by the model's native computational pathway, not by a post-hoc surrogate. The prototype layer is a bottleneck that forces the model to make predictions based on prototype similarity. This architectural constraint ensures that the similarity scores used for explanation are the actual values the model used for its decision, guaranteeing a high degree of faithfulness and eliminating the introduction-to-explanation gap.

The method naturally supports 'why this, not that?' reasoning. By examining the similarity scores to prototypes of other classes, a user can understand why a different outcome was not predicted. The explanation shows not only the closest matching prototype for the predicted class but also the distance to the nearest prototype of a contrasting class, providing a decision boundary perspective.

The complete set of learned prototypes serves as a global summary of the model's knowledge. By visualizing all prototypes, a domain expert can audit the entire model's learned concepts at once. This enables a qualitative check: do the prototypes represent physically meaningful states or spurious correlations? This global view is a powerful tool for model validation and debugging before deployment.