🤖 machine learning

INSIGHTS: Demonstration-Based Summaries of Time Series Predictors

This paper introduces INSIGHTS, a model-agnostic, user-centric framework that generates diverse, demonstration-based global summaries of time series models to provide domain experts with a stable and comprehensive understanding of model behavior, addressing the current gap in global explainability for time-series data.

Original authors: Bar Eini Porat, Rom Gutman, Uri Shalit, Ofra Amir

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bar Eini Porat, Rom Gutman, Uri Shalit, Ofra Amir

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart computer that predicts the future based on a massive, endless stream of data—like stock prices, heart rates, or weather patterns. This computer is a "black box." It gives you answers, but it doesn't tell you how it thinks or what it pays attention to.

Currently, if you want to understand this computer, you have to look at one specific moment in time (like asking, "Why did the stock drop right now?"). But that's like trying to understand a whole movie by only watching a single frame. You miss the plot, the character arcs, and the big dramatic moments.

The paper introduces INSIGHTS, a new tool designed to give you the "movie trailer" instead of just a single frame. Here is how it works, using simple analogies:

1. The Problem: Too Much Data, Not Enough Time

Imagine a doctor trying to understand a patient's heart monitor. The machine records data every second for weeks. If the doctor tried to read every single second, they would never sleep. They need a summary.

Existing tools are like a librarian who hands you a random page from a book. Sometimes it's interesting, sometimes it's boring, and often it misses the main plot points. Other tools try to summarize the whole book but end up being too complicated or only work for specific types of books (models).

2. The Solution: INSIGHTS (The "Executive Summary" Generator)

INSIGHTS is a tool that picks a small, manageable set of examples (a "Time Series Summary") to show you how the computer model behaves. It's like an Executive Summary for a business report: instead of reading 500 pages of data, you read 10 carefully chosen pages that tell you everything important.

It does this by balancing two things:

Importance: It looks for the "drama." In a hospital, this might be a sudden spike in blood pressure. In stocks, it might be a massive crash. It ignores the boring, routine parts.
Diversity: It makes sure it doesn't just show you 10 different versions of the same "drama." It ensures you see a mix of different types of events (a slow rise, a sudden drop, a weird oscillation) so you get the full picture.

3. How It Works: The "Taste Tester" Analogy

Imagine you are a chef trying to figure out if a new soup recipe is good. You can't taste the whole pot.

Old Methods: Might just taste the soup from the top of the pot (random) or taste only the spicy parts (biased).
INSIGHTS: Uses a set of "taste rules" (called Utility Functions) that you define.
- Rule 1: "If the temperature changes too fast, that's important."
- Rule 2: "If the salt level goes outside the normal range, that's important."
- Rule 3: "If the trend is going up, that's important."

The tool scans the whole pot, finds the moments that break these rules, and then picks a diverse group of them. It ensures you get a sample that includes a "too hot" moment, a "too salty" moment, and a "strange swirl" moment, giving you a complete understanding of the soup's behavior without tasting every drop.

4. What They Found (The Results)

The researchers tested this tool in three ways:

The "Event Capture" Test: They fed the tool data with hidden "events" (like sudden spikes). INSIGHTS found almost 100% of these events, while other tools missed many of them. It was also much faster and used less computer memory (like a sports car vs. a heavy truck).
The "Doctor" Test: They showed summaries to Intensive Care Unit (ICU) doctors.
- When doctors saw the INSIGHTS summary, they could spot more different types of model behaviors (like how the model handles sudden changes vs. slow trends).
- They found the INSIGHTS examples were more useful and less repetitive than those from other tools.
- Doctors said these summaries would be great for shift changes, helping one doctor quickly understand what happened to a patient during the previous shift.
The "Student" Test: They asked students to guess how a stock-predicting model worked based on summaries.
- Students who saw the INSIGHTS summaries understood the model slightly better, especially on the harder questions.
- They felt less "mentally tired" (cognitive load) and trusted the explanation more.

5. The Bottom Line

INSIGHTS is a practical, efficient way to turn a mountain of time-series data into a small, diverse, and meaningful story. It doesn't just show you random data; it shows you the key moments that define how a model thinks.

It is designed to be a "front door" for understanding complex models: you look at the summary first to get the big picture, and then you can dive deeper into specific examples if you need to. It works for any type of model (model-agnostic) and is fast enough to use on huge datasets where other tools would crash.

Problem Statement
While explainability methods for time-series models have advanced rapidly in the realm of local, instance-level attributions (e.g., SHAP, LIME, shapelets), global explanations remain underdeveloped. Existing global approaches are often model-specific or modality-specific, failing to address scenarios where diverse model families coexist. Furthermore, current methods struggle to identify which time-series segments are critical for inspection, particularly when "importance" is domain-dependent and driven by rare or unlabeled events (e.g., sudden blood pressure spikes in an ICU or stock market surges). There is a lack of reliable, model-agnostic frameworks to construct global summaries that guide stakeholders in understanding a model's behavior across different regimes and critical events.

Methodology: INSIGHTS
The paper introduces INSIGHTS, a model-agnostic, user-centric algorithm designed to generate global explanations for time-series prediction models through "Time-Series Summaries" (TSS). A TSS is defined as a concise, manageable collection of samples (segments or events) that highlight key model behaviors.

The algorithm operates on two primary criteria: Importance and Diversity.

Utility Functions (Importance): INSIGHTS utilizes user-defined utility functions to rank time-series segments based on domain-specific significance. The paper proposes three default utility functions adapted from prior work:
- Overall Trend: Captures increasing or decreasing slopes.
- Exceeding Normal Range: Identifies values that deviate from domain-relevant norms (using sigmoid functions).
- Sudden Changes Trend Deviation: Detects abrupt deviations from expected trends.
  These functions can operate on raw values or user-defined feature maps (e.g., frequency-domain features).
Diversity Selection: The algorithm iteratively selects samples to maximize behavioral diversity. It employs a greedy selection process where, in each round, it proposes top-importance candidates from a specific utility perspective and selects the one that maximizes the distance from the currently selected set.
- INSIGHTS-TW: Uses Dynamic Time Warping (DTW) to calculate pairwise distances, accounting for temporal context and scale differences.
- INSIGHTS-critic: An alternative using a non-time-series-specific diversity calculation (based on MMD-Critic prototypes).
Algorithm Flow: The process initializes by sorting samples into buckets based on utility scores. It iteratively selects candidates that satisfy overlap constraints and maximize diversity (measured via DTW distance or prototype distance) until a target summary size $m$ is reached. It also allows for the inclusion of median-importance prototypes.

Key Contributions

Model-Agnostic Global Summaries: INSIGHTS provides a framework for generating global explanations for time-series prediction that is not tied to specific model architectures (e.g., deep learning vs. regression).
Balanced Selection: It uniquely balances the importance of samples (via utility functions) with the diversity of behaviors (via time-series-aware distance metrics), ensuring the summary covers distinct regimes and critical events rather than just the most frequent patterns.
Efficiency: The method is computationally and memory efficient. While baseline methods like ProtoDash and MMD-Critic exhibit quadratic complexity and high memory usage, INSIGHTS operates with log-linear complexity relative to input size, making it scalable for large physiological or financial datasets.
User-Centric Design: The approach prioritizes simplicity and transparency, allowing stakeholders to audit carefully chosen examples rather than arbitrary datasets.

Results
The paper evaluates INSIGHTS through three complementary studies:

Event Capture (Computational): On synthetic data and the PhysioNet Sleep-EDF EEG dataset, INSIGHTS-TW achieved 100% event-type coverage and 100% event example coverage across various summary sizes. In contrast, baselines like MMD-Critic showed low coverage, and ProtoDash failed to complete on the large EEG dataset due to resource exhaustion. INSIGHTS demonstrated linear runtime growth compared to the quadratic growth of baselines.
Expert Study (ICU Clinicians): Semi-structured interviews with ICU clinicians revealed that INSIGHTS summaries prompted broader insights (oscillations, trend adaptation, sudden events) compared to ProtoDash (conservatism) or random sampling. Clinicians found INSIGHTS examples highly informative and noted their utility in model selection and shift-change handovers.
User Study (Stock Prediction): In a task where participants identified model behavior based on summaries, the INSIGHTS group showed slightly higher accuracy (0.77 vs. 0.68) than the ProtoDash group, though the difference was marginally significant ( $p=0.08$ ). However, INSIGHTS summaries were rated significantly higher in terms of completeness, lower cognitive load, and perceived model reliability. Graduate students (simulating semi-experts) benefited most from the diverse INSIGHTS summaries.

Significance and Claims
The paper claims that INSIGHTS effectively constructs comprehensive, diverse subsets of time-series data that are manageable for human evaluation. It asserts that the method is preferred by domain experts for providing a stable understanding of model behavior and identifying high-quality samples. By offering a principled set of segments, INSIGHTS enables the scalable application of local, model-agnostic explainers to global time-series problems. The authors position the method as a practical, efficient front-end for interactive explanation systems, particularly in domains where rare events and temporal context are critical. The paper remains modest, acknowledging limitations such as the reliance on greedy selection (which does not guarantee global optimality) and the current focus on univariate series, noting that multivariate extensions are a natural next step.

1. The Problem: Too Much Data, Not Enough Time

2. The Solution: INSIGHTS (The "Executive Summary" Generator)

3. How It Works: The "Taste Tester" Analogy

4. What They Found (The Results)

5. The Bottom Line

More like this