ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting

TL;DR

ProtoTS presents a hierarchical prototypical framework for explainable time series forecasting with exogenous variables, balancing accuracy and interpretability. It introduces a multi-channel embedding and bottleneck fusion to compute prototype similarities, and organizes prototypes into a root–child hierarchy to capture global and local temporal patterns. Empirical results on LOF and EPF show state-of-the-art forecast accuracy and improved interpretability, with case studies illustrating expert steerability through prototype editing. The approach offers a practical pathway to global explanations and actionable model adjustments in high-stakes forecasting tasks.

Abstract

While deep learning has achieved impressive performance in time series forecasting, it becomes increasingly crucial to understand its decision-making process for building trust in high-stakes scenarios. Existing interpretable models often provide only local and partial explanations, lacking the capability to reveal how heterogeneous and interacting input variables jointly shape the overall temporal patterns in the forecast curve. We propose ProtoTS, a novel interpretable forecasting framework that achieves both high accuracy and transparent decision-making through modeling prototypical temporal patterns. ProtoTS computes instance-prototype similarity based on a denoised representation that preserves abundant heterogeneous information. The prototypes are organized hierarchically to capture global temporal patterns with coarse prototypes while capturing finer-grained local variations with detailed prototypes, enabling expert steering and multi-level interpretability. Experiments on multiple realistic benchmarks, including a newly released LOF dataset, show that ProtoTS not only exceeds existing methods in forecast accuracy but also delivers expert-steerable interpretations for better model understanding and decision support.

Figures (6)

• Figure 1: (a) An example of time series forecasting with numerous heterogeneous variables, where exogenous variables (e.g., temperature, holiday) influence the evolution of endogenous variables (e.g., electric load). (b) Illustration of prototypical explanation method: a set of learned prototypes provides a user-friendly global overview of typical temporal patterns. For each instance, model computes its similarity to all prototypes to form a prediction, enabling detailed local interpretation.
• Figure 2: The overall framework of ProtoTS, which comprises two main modules: the multi-channel prototype similarity computation module and the hierarchical prototype learning module.
• Figure 3: Sensitivity analysis: (a) Effect of increasing root prototype numbers from {6, 9, 12, 15, 18}. (b) Data efficiency as the training data proportion increases from 50% to 100%.
• Figure 4: Case study of interpretability and steerability: (a) Activation patterns of prototypes, which each colored bar shows the degree of activation of its corresponding prototype each day. This visualization shows seasonal, weekly, and holiday-related prototype activations. (b) The learned prototype hierarchy that illustrates typical temporal patterns and their correlations with covariates for interpretable load forecasting. After the expert splits the first prototype at layer-2, the mean MSE across five random seeds decreases from 0.099 to 0.090 for Spring Festival related predictions.
• Figure 5: User Study Questionnaire. Bottom: An overview of the questionnaire's purpose, question formats, and response instructions. Left: The Variable-related Questionnaire, where users are asked to infer input variables based on the provided explanations. Each question has a single correct answer, used to evaluate the accuracy of the system's explanations. Right: The System Usability Scale (SUS) Questionnaire, which captures users' subjective assessments to measure the usability of the current system.