Explainable classification of astronomical uncertain time series

This paper proposes an uncertainty-aware, explainable-by-design subsequence-based model that achieves state-of-the-art classification performance for astronomical uncertain time series by incorporating data uncertainty as an input, thereby enabling domain experts to inspect predictions and potentially inspire new theoretical astrophysics developments.

The Gist

Imagine you are a detective trying to identify different types of birds based on their songs. But here's the catch: you are listening through a thick fog, and every time a bird chirps, you aren't 100% sure if you heard a "tweet" or a "chirp." Sometimes the sound is fuzzy, and your recording device adds a little bit of static.

Most computer programs (Machine Learning) are like detectives who ignore the fog. They just guess the song based on the clearest part of the recording. If the recording is too fuzzy, they get confused and make mistakes.

This paper introduces a new kind of detective: one who knows the fog is there and uses it to help solve the case.

Here is the breakdown of what the authors did, using simple analogies:

1. The Problem: The "Fuzzy" Light Curves

In astronomy, scientists watch stars and galaxies to see how their brightness changes over time. This change is called a Light Curve.

• The Reality: When we look at distant stars, our measurements aren't perfect. There is "noise" from clouds, the moon, or the telescope itself.
• The Data: Instead of a single line showing the brightness, the data comes with a "fuzzy zone" (an error bar). It's like saying, "The star was probably this bright, but it could be a little brighter or a little dimmer."
• The Challenge: Old computer methods tried to ignore this fuzziness. They tried to draw a perfect line through the mess. But because the data is so messy and the different types of stars look very similar, these old methods failed to tell them apart.

2. The Solution: The "Pattern Hunter" (uSAST)

The authors created a new method called uSAST (Uncertain Scalable and Accurate Subsequence Transform).

Think of a Light Curve as a long, winding road.

• Old Method: Tries to look at the entire road at once to guess where it goes. If the road is foggy, the whole picture is blurry.
• New Method (uSAST): Instead of looking at the whole road, it cuts the road into small, manageable segments (subsequences). It looks for specific "landmarks" or patterns in these small segments.
  • Example: "Ah, this specific 10-second bump in the road looks exactly like a 'Supernova'!"
  • The Twist: This new detective doesn't just look at the shape of the bump; it also looks at how foggy the bump is. If a bump is very fuzzy, the detective knows, "Okay, this is a tricky one, but I can still use the fuzziness to help me decide."

3. Why is it "Explainable"? (The "Why" Matters)

Most powerful AI models are "Black Boxes." You feed them data, and they give an answer, but they can't tell you why. It's like a wizard saying, "I know it's a supernova because the magic numbers say so."

This new method is "Explainable by Design."

• It's like a detective who points to the specific clue on the board and says, "I know this is a Supernova because of this specific spike in brightness at this specific time."
• Because the model highlights these specific patterns, real human astronomers can look at them and say, "Yes, that matches what we expect from a Supernova!" This builds trust.

4. The Results: A Win for Science

The authors tested their method on a real dataset from the PLAsTiCC challenge (a huge competition to classify stars).

• The Competition: They compared their method against the best existing AI tools.
• The Outcome: Their method was just as accurate as the top "Black Box" models (getting about 70% right), but with a huge advantage: it could explain its answers.
• The "Fog" Factor: They proved that by actually using the "fuzziness" (uncertainty) in the data, the model got smarter. If they ignored the fuzziness, the model got dumber.

5. The Big Picture: Why Should We Care?

Imagine the Vera C. Rubin Observatory, a giant new telescope that will take pictures of the entire sky every night. It will generate so much data that humans can't look at it all.

• We need AI to sort through billions of stars.
• But we can't just trust a "Black Box" AI to tell us about the history of the universe. We need to know why it thinks a star is a supernova.
• This paper gives us a tool that is both smart (accurate) and honest (explainable). It helps astronomers find new types of stars and understand the universe, even when the data is messy.

Summary Analogy

Imagine trying to identify a song by listening to a radio with bad reception.

• Old AI: Tries to guess the song by ignoring the static. It often guesses wrong because the static hides the melody.
• New AI (uSAST): Listens to the melody and the static. It realizes, "The static is heavy here, which means this part of the song is likely a drum beat." It uses the noise to help identify the song.
• The Bonus: When it guesses, it points to the exact part of the recording and says, "I heard a drum beat here, and the static confirmed it."

This paper is about teaching computers to be better listeners in a noisy world, so we can understand the universe a little better.

Technical Summary

Here is a detailed technical summary of the paper "Explainable Classification of Astronomical Uncertain Time Series" by Mbouopda et al.

1. Problem Statement

The paper addresses the challenge of classifying Uncertain Time Series (uTS), specifically within the domain of astrophysics.

• Nature of Data: Unlike standard time series where values are precise real numbers, uTS consists of pairs $(v, \delta)$, where $v$ is the best estimate and $\delta$ is the measurement error (uncertainty). In astronomy, these represent light curves (brightness over time) where flux measurements have inherent noise due to atmospheric conditions, moonlight, or instrument limitations.
• The Gap: While black-box machine learning models (e.g., deep learning, random forests) achieve high accuracy, they lack explainability, which is critical for scientific discovery. Existing interpretable methods (like Shapelet Transform) generally fail on uTS data because they either ignore uncertainty or cannot handle the specific noise structures of astronomical data.
• Specific Context: The study focuses on the PLAsTiCC dataset, a real-world benchmark containing multivariate uncertain time series representing 14 types of astronomical transients (e.g., Supernovae, Kilonovae, Active Galactic Nuclei). The dataset is highly imbalanced and contains missing values.

2. Methodology: uSAST (Uncertain Scalable and Accurate Subsequence Transform)

The authors propose uSAST, a method designed to be "explainable-by-design" while explicitly propagating data uncertainty through the classification pipeline.

Core Concepts

• Uncertain Subsequence Transform: The method transforms the raw time series into a feature vector based on distances to a set of representative subsequences (patterns).
• Uncertainty Propagation: Unlike Conformal Learning (which estimates uncertainty in the output), uSAST treats uncertainty as an input feature. It uses the Uncertain Euclidean Distance (UED) to calculate distances between time series and subsequences, ensuring the classifier learns decision boundaries that account for the error bars in the data.
• $\epsilon$-Similarity and Frequency: To handle redundancy in repetitive patterns (common in light curves), the authors introduce an $\epsilon$-similarity threshold. Subsequences within this distance are considered similar. The method also counts the frequency of pattern occurrences, not just the minimum distance.

Algorithm Steps (SAST+ and uSAST+)

1. Reference Selection: Randomly select $k$ reference instances per class (in experiments, $k=1$).
2. Pattern Generation: Generate all possible subsequences from the reference instances.
3. Deduplication: Apply $\epsilon$-similarity to remove redundant subsequences, creating a compact dictionary of unique patterns.
4. Transformation: For every time series in the dataset, compute a feature vector containing:
   • The distance to each unique pattern (using UED for uSAST).
   • The frequency count of how many times each pattern appears in the series.
   • The uncertainty values associated with these distances.
5. Classification: Train a supervised classifier (e.g., XGBoost, Random Forest) on the transformed feature matrix. Since standard classifiers cannot natively handle "uncertain" numbers, the uncertainty is represented as a separate feature (e.g., mean and standard deviation) alongside the value.

3. Key Contributions

• Explainable-by-Design: The model relies on subsequences (shapelets) that are directly interpretable by domain experts. Experts can inspect the specific time intervals and wavelengths that drove a classification decision.
• Native Uncertainty Handling: The method is the first to successfully apply subsequence-based classification to real-world uTS data by integrating uncertainty directly into the distance metric and feature extraction, rather than treating it as a post-hoc estimation.
• Open Science: The authors released the preprocessed PLAsTiCC dataset, the source code, and the full experimental results, addressing a lack of reproducibility in previous uTS research.
• Ablation of Uncertainty: The study explicitly demonstrates that ignoring uncertainty leads to significantly lower performance, validating the necessity of uncertainty-aware models in this domain.

4. Experimental Results

The method was evaluated on the PLAsTiCC training set (7,848 objects, 15 classes) using 80/20 train/test splits.

• Performance vs. Baselines:
  • Shapelet-based methods (STC, UST): Failed completely (F1 $\approx$ 0) because they could not find valid discriminative shapelets in the noisy, imbalanced data.
  • SOTA Black-box Methods: ROCKET achieved the highest F1 score (~0.75) but is not explainable. MUSE and XEM achieved ~0.71 and ~0.69 respectively.
  • uSAST: Achieved an F1 score of ~0.70 (specifically 0.70 $\pm$ 0.01 with XGBoost), comparable to SOTA methods.
• Impact of Uncertainty:
  • Comparing SAST (ignoring uncertainty) vs. uSAST (using uncertainty): uSAST showed a 6% gain in precision, 5% in recall, and 6% in F1 score.
  • The model became more confident (lower LogLoss) when uncertainty was included.
• Astronomical Insights:
  • The model performed significantly better on Galactic objects (F1 0.94) than Extragalactic ones (F1 0.67), reflecting the higher signal-to-noise ratio and cyclic nature of galactic variables.
  • It successfully distinguished Type Ia Supernovae (SNIa), a critical class for cosmology, achieving an F1 of 0.82 in a binary classification task against other classes.
• Explainability Validation:
  • The extracted discriminative subsequences matched physical expectations (e.g., identifying the "plateau" phase in Type II Supernovae or the dip in Eclipsing Binaries).
  • The model correctly identified that certain classes were distinguishable primarily by their uncertainty levels (e.g., distant objects having larger error bars), a feature black-box models often miss.

5. Significance and Conclusion

This work bridges the gap between high-performance machine learning and scientific interpretability in astrophysics.

• Scientific Trust: By providing "faithful" explanations, the method allows astronomers to trust the model's predictions and potentially discover new physical patterns in the data (e.g., unexpected light curve shapes).
• Future of LSST: As the Vera C. Rubin Observatory begins operations, it will generate massive amounts of uncertain photometric data. uSAST offers a scalable, interpretable framework to process this data without sacrificing accuracy.
• Generalizability: While tested on astronomy, the methodology is applicable to any domain involving uncertain time series (e.g., medical monitoring, financial forecasting with noisy sensors).

Limitations: The primary limitation is computational complexity. The method is slower than black-box SOTA methods (taking ~43 hours vs. minutes for ROCKET), though the authors argue this is a necessary trade-off for interpretability and uncertainty handling. Future work aims to optimize the subsequence selection to reduce runtime.