ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting

ProtoTS is a novel interpretable time series forecasting framework that achieves high accuracy and transparent decision-making by modeling hierarchical prototypical patterns to capture both global trends and local variations, thereby enabling expert-steerable explanations for high-stakes scenarios.

The Gist

Imagine you are trying to predict the weather for the next week. You have a super-smart computer (a Deep Learning model) that looks at thousands of data points: temperature, humidity, wind speed, historical patterns, and even the day of the week. The computer gives you a perfect forecast, but when you ask, "Why did you predict rain on Tuesday?" it just says, "Because the math says so."

In high-stakes jobs like managing a power grid or predicting stock prices, "because the math says so" isn't good enough. You need to know why so you can trust the decision or fix it if it's wrong.

This paper introduces ProtoTS, a new way to make time-series forecasting both super accurate and easy to understand. Here is how it works, using some simple analogies.

1. The Problem: The "Black Box" and the "Partial Puzzle"

Current AI models are like Black Boxes. They give great answers, but you can't see inside to see how they thought.
Other "explainable" models are like Puzzle Solvers who only show you one piece of the puzzle at a time. They might tell you, "Temperature caused the spike at 2 PM," but they fail to explain the whole story: "Why does the load curve look like a mountain range with three peaks every day, and how does the holiday season change that shape?"

2. The Solution: The "Pattern Library" (Prototypes)

ProtoTS changes the game by teaching the AI to learn Prototypes.

Think of a Prototype as a "Template" or a "Classic Recipe."

• Instead of memorizing every single day's weather, the AI learns a few "Classic Days."
• Prototype A: "A typical hot summer weekday." (High AC usage, peak at 2 PM).
• Prototype B: "A typical rainy holiday." (Low activity, flat line).
• Prototype C: "A typical winter workday." (Morning and evening peaks).

When the AI needs to make a prediction, it doesn't just guess. It looks at the current situation and asks: "Which of my Classic Recipes does this look most like?"

3. The Secret Sauce: The "Tree of Knowledge" (Hierarchical Learning)

The paper's biggest innovation is that these "Classic Recipes" are organized in a Family Tree.

• The Root (The Grandparents): These are the big, broad patterns.
  • Example: "Workday" vs. "Weekend" vs. "Holiday."
  • This gives you a Global View. You can instantly see, "Oh, the AI is predicting a 'Holiday' pattern today."
• The Branches (The Parents): These split the big patterns into slightly more specific ones.
  • Example: Under "Holiday," we have "Long Holiday" (like Spring Festival) and "Short Holiday" (like a long weekend).
• The Leaves (The Children): These are the tiny, detailed variations.
  • Example: Under "Long Holiday," we have "Spring Festival 2024" (where people stay home all day) vs. "Spring Festival 2023" (where people traveled more).

Why is this cool?
It allows for Expert Steering. If you are an energy expert, you can look at the "Tree" and say, "Hey, the 'Spring Festival' branch is too simple. Let's split it into 'Pre-Holiday' and 'During-Holiday'." The AI listens, creates a new branch, and instantly becomes more accurate and more explainable.

4. Handling the Chaos: The "Multi-Channel Kitchen"

Real-world data is messy. You have numbers (temperature), categories (is it a holiday?), and time (what time of day?).
ProtoTS uses a Multi-Channel Kitchen:

• It puts the "Temperature" ingredients in one bowl.
• It puts the "Holiday" ingredients in another bowl.
• It mixes them carefully using a special "Bottleneck" filter (like a sieve) to remove the junk (noise) and keep only the tasty, important flavors.
• Finally, it matches this clean mixture against its "Pattern Library" to make the prediction.

5. The Results: Smarter and Safer

The authors tested ProtoTS on real-world data, like predicting electricity loads for a whole country.

• Accuracy: It beat all the other top AI models. It reduced errors by nearly 50% in some cases.
• Trust: They showed the results to human experts. The experts could look at the "Pattern Tree" and immediately understand why the AI made a prediction.
• Control: In a test, an expert manually tweaked a "Spring Festival" pattern. The AI's prediction for that specific time got even better. It's like a pilot and a co-pilot working together, rather than a pilot flying blind.

Summary

ProtoTS is like giving the AI a Lego set of time patterns instead of a magic black box.

1. It builds a Library of Templates (Prototypes) for different types of days.
2. It organizes them in a Tree so you can see the big picture and the small details.
3. It lets Humans step in, look at the tree, and say, "Let's refine this branch," making the AI smarter and more trustworthy.

It's the difference between a calculator that just gives you a number and a financial advisor who can show you the chart, explain the trend, and let you adjust the strategy.

Technical Summary

1. Problem Statement

Time series forecasting is critical in high-stakes domains like energy management and load forecasting. While deep learning models achieve high accuracy, they often function as "black boxes," lacking transparency. Existing interpretable methods suffer from two main limitations:

• Output-side limitation (C1): Current methods typically explain predictions at individual time steps (local explanations) but fail to provide a global overview of the overall temporal patterns (e.g., explaining why a forecast curve has three specific peaks throughout a day).
• Input-side limitation (C2): Explanations often focus only on endogenous variables (the target history) or specific variable types, failing to capture the complex interactions between heterogeneous exogenous variables (e.g., how temperature, holidays, and day-of-week jointly influence the forecast).

The core challenge is balancing predictive accuracy with multi-level interpretability and expert steerability in the presence of heterogeneous data.

2. Methodology: ProtoTS Framework

ProtoTS is an end-to-end framework that models time series forecasting as a matching process between input instances and a set of learned prototypical temporal patterns. The framework consists of two primary modules:

A. Multi-Channel Prototype Similarity Computation

This module handles the encoding of heterogeneous inputs and computes the similarity between an input instance and the prototypes.

• Multi-Channel Embedding:
  • Endogenous variables (e.g., past load) are encoded via a non-linear projection (MLP).
  • Exogenous variables are split into discrete (e.g., holidays, day of week) and continuous (e.g., temperature) channels. Discrete variables use embedding tables, while continuous variables use variable-specific non-linear projections.
• Bottleneck Channel Fusion:
  • The embeddings from all channels are aggregated (summed) and passed through a bottleneck layer (using an MLP-Mixer architecture).
  • This mechanism filters out irrelevant noise from the heterogeneous inputs while preserving predictive components, projecting high-dimensional data into a lower-dimensional latent space and back.
• Prototype Similarity:
  • The model computes the similarity between the aggregated instance embedding ($\hat{Z}$) and a set of learnable prototype embeddings ($\mu$).
  • The similarity score (via softmax over Euclidean distance) determines the weight of each prototype in the final prediction.

B. Hierarchical Prototype Learning

To address the trade-off between global interpretability and local accuracy, ProtoTS organizes prototypes in a tree structure:

• Root Level (Global): A small set of coarse prototypes captures broad, recurring temporal patterns (e.g., "Summer," "Workday," "Holiday"). These provide a high-level, human-understandable summary of the model's logic.
• Child Level (Local): Each root prototype can be recursively split into child prototypes to capture fine-grained variations (e.g., splitting "Holiday" into "Spring Festival" vs. "Normal Holiday").
• Prediction Generation: The final forecast is a weighted combination of the temporal patterns ($p$) of all active prototypes, modulated by both root-level and child-level similarity scores.
• Expert Steering: The hierarchy allows experts to intervene. If a specific pattern is poorly represented, an expert can manually split a prototype or edit its temporal pattern to incorporate domain knowledge.

3. Key Contributions

1. Hierarchical Prototype Learning Strategy: A novel approach that decouples global pattern understanding (root prototypes) from local detail modeling (child prototypes), enabling both high accuracy and multi-level interpretability.
2. Multi-Channel Similarity Computation: A mechanism specifically designed to model interactions between heterogeneous exogenous variables (discrete/continuous) and endogenous history, filtering noise via a bottleneck fusion layer.
3. Expert-Steerable Interpretability: Unlike post-hoc explanation methods, ProtoTS is inherently interpretable. Experts can directly inspect, split, or edit prototypes to refine model behavior, bridging the gap between data-driven learning and domain expertise.
4. Holistic Temporal Pattern Decoding: Unlike previous prototype methods that map to single outputs (classification), ProtoTS decodes prototypes into sequences of regression results (forecast curves), explaining the entire temporal evolution.

4. Experimental Results

The authors evaluated ProtoTS on multiple benchmarks, including a newly released LOF (Load Forecasting) dataset and the EPF (Electricity Price Forecasting) dataset.

• Forecasting Accuracy:
  • On the LOF dataset, ProtoTS achieved state-of-the-art performance, reducing MSE by 48.3% and MAE by 20.9% compared to the best baseline (TimeXer).
  • On the EPF dataset, it reduced MSE and MAE by 8% across five markets.
  • It demonstrated superior data efficiency, maintaining stable performance even when training data was reduced to 50%, whereas baselines degraded significantly.
• Interpretability Evaluation:
  • User Precision (UPrec): In a user study, ProtoTS achieved 77% accuracy in helping users infer input variables from explanations, significantly outperforming TFT (64%) and N-BEATSx (62%).
  • System Usability Scale (SUS): ProtoTS scored 73.36, indicating high usability compared to baselines (TFT: 29.74).
• Case Study (Steerability):
  • In a load forecasting case study, experts identified that the "Spring Festival" prototype was too generic. By manually splitting it into "Base Spring Festival" and "Special Spring Festival" patterns, the model's MSE for that period dropped by 0.009, demonstrating that expert intervention directly improves accuracy.

5. Significance

ProtoTS represents a significant shift in time series forecasting by moving away from "black-box" accuracy toward transparent, steerable intelligence.

• Trust in High-Stakes Scenarios: It provides the necessary transparency for domains like power grid management, where understanding why a peak is predicted is as important as the prediction itself.
• Human-in-the-Loop: The framework uniquely enables domain experts to refine models through prototype editing, making it a collaborative tool rather than just an automated predictor.
• Handling Heterogeneity: It effectively solves the challenge of integrating diverse data types (weather, calendar, historical load) without sacrificing interpretability, a common failure point in current Transformer-based models.

In conclusion, ProtoTS successfully demonstrates that high-accuracy forecasting and deep interpretability are not mutually exclusive, offering a scalable framework for explainable AI in time series.