Supernova scores for active anomaly detection

This paper presents a hybrid active anomaly detection framework that integrates a supervised supernova probability score into the PineForest algorithm to significantly enhance the discovery efficiency of supernovae and other rare transients in large-scale time-domain surveys like the Zwicky Transient Facility.

The Gist

Imagine you are a librarian in a library that receives millions of new books every night. Most of these books are either blank pages, scribbled notes, or boring copies of the same old story. But hidden inside this mountain of paper are a few dozen masterpieces—rare, brilliant stories that could change history.

Your job is to find those masterpieces before the library collapses under the weight of the junk.

This is exactly the challenge astronomers face with modern sky surveys like the Zwicky Transient Facility (ZTF). Every night, telescopes take pictures of the sky, generating millions of "light curves" (graphs showing how bright an object gets over time). Most of these are just stars blinking normally, dust, or camera glitches. But hidden in there are Supernovae (exploding stars), which are the "masterpieces" astronomers want to study.

Here is how the authors of this paper solved the problem, using a mix of smart computers and human intuition.

1. The Problem: The Needle in a Haystack

If you try to find a needle in a haystack by looking at every single piece of hay, you'll go crazy.

• Supervised Learning (The "Expert" approach): You can teach a computer to recognize a supernova by showing it thousands of examples. It gets very good at spotting known supernovae. But if a new, weird type of explosion happens, the computer might ignore it because it doesn't look like the examples it was taught.
• Unsupervised Learning (The "Curious" approach): You can tell the computer, "Show me anything weird." This finds all kinds of strange things, but it's like a metal detector that beeps at every soda can, bottle cap, and coin. It's too noisy to find the specific treasure you want.

2. The Solution: The "Hybrid" Strategy

The authors created a hybrid strategy that combines the best of both worlds. Think of it as hiring a detective who has a specific "Wanted" poster, but is also smart enough to notice if the criminal is wearing a disguise.

Step A: The "Supernova Score" (The Wanted Poster)

First, they trained a computer model to act like a Supernova Detective. They fed it data from confirmed supernovae and taught it to give every object in the sky a "Supernova Score" (or SN-score).

• If an object looks like a supernova, it gets a high score (9/10).
• If it looks like a normal star or a glitch, it gets a low score (1/10).

Step B: The "PineForest" (The Smart Search)

Next, they used a tool called PineForest. Imagine this as a smart search engine that doesn't just look at the "Wanted Poster." Instead, it looks at the entire library and asks: "What is the strangest thing here?"

Here is the magic trick: They gave the search engine the Supernova Score as an extra clue.

• Without the clue: The search engine might get distracted by a weirdly colored star that isn't a supernova.
• With the clue: The search engine knows, "Hey, this object has a high Supernova Score and it's acting strangely. Let's look at that one first!"

Step C: The Human Guide (The "Priors")

The system also lets a human expert say, "I know what a real supernova looks like; here are 10 examples." The computer uses these examples to "prune" its search, cutting away the branches of the tree that don't match the expert's intuition. This makes the search much faster.

3. The Results: Finding Hidden Treasures

By using this combined method, the team looked at ten specific patches of the sky and found:

• 7 New Supernovae: Stars that exploded but were missed by previous automated systems.
• 1 Active Galaxy: A galaxy with a supermassive black hole that was acting up.
• 1 Weird Star: A star in our own galaxy (SNAD283) that is glowing with helium and behaving strangely—neither a normal nova nor a dwarf nova.
• 2 "Sibling" Supernovae: Two galaxies where two different stars exploded in the same place, years apart. This is like finding two different actors playing the same role in the same theater, years apart.

Why This Matters

This method is like upgrading from a flashlight to a metal detector with a GPS.

• It doesn't just find what we already know (like the flashlight).
• It doesn't just find anything weird (like the metal detector).
• It finds the specific things we care about (supernovae) while still keeping an eye out for other surprises.

The authors say this is crucial for the future, especially for the Vera C. Rubin Observatory, which will soon take pictures of the entire sky every few nights. Without a smart system like this, the data would be too overwhelming for humans to ever find the most exciting discoveries.

In short: They taught a computer to recognize the "sound" of an exploding star, then used that knowledge to help a smart search algorithm find the quietest, rarest explosions in the universe, all while keeping an open mind for other cosmic mysteries.

Technical Summary

Here is a detailed technical summary of the paper "Supernova scores for active anomaly detection."

1. Problem Statement

Time-domain astronomy surveys, such as the Zwicky Transient Facility (ZTF) and the upcoming Vera C. Rubin Observatory (LSST), generate massive volumes of data (hundreds of millions of light curves) containing a vast majority of routine variability and instrumental artifacts.

• The Challenge: Scientifically valuable transients (like Supernovae, SNe) are extremely rare compared to the background noise.
• Limitations of Existing Methods:
  • Supervised Learning: While effective at filtering known classes, these models struggle with extreme class imbalance and often fail to detect rare, novel, or unlabelled events.
  • Unsupervised Anomaly Detection: These methods offer broad discovery potential but lack sensitivity to specific classes of interest (e.g., SNe) unless explicitly guided, often resulting in low discovery efficiency for targeted searches.
• Goal: Develop a hybrid strategy that leverages supervised learning to guide unsupervised anomaly detection, thereby accelerating the discovery of specific rare events (SNe) without sacrificing the ability to find other diverse astrophysical anomalies.

2. Methodology

The authors propose a hybrid framework integrating a supervised binary classifier with the PineForest active anomaly detection algorithm.

A. Data Preparation

• Source Data: ZTF Data Release 23 (DR23), specifically $r$-band light curves.
• Training Sample (Positive Class): 674 spectroscopically confirmed SNe from the ZTF Bright Transient Survey (BTS) cross-matched with DR23.
• Negative Class: A combination of 6,300 non-SN objects from the Anomaly Knowledge Base (AKB) and 10,000 randomly selected objects from the main dataset.
• Feature Extraction: 47 astronomical features were extracted from light curves using the light-curve Python library (e.g., amplitude, skewness, periodogram peaks, $\chi^2$ of fits).
• Dataset Scope: The analysis focused on 10 extragalactic fields (outside the Galactic plane) containing approximately 720 million light curves.

B. The Hybrid Pipeline

1. Supervised Binary Classifier (SN-score):

   • Model: Random Forest classifier.
   • Training Strategy: To handle class imbalance, the model was trained with a 1:1000 class weight ratio (SNe vs. non-SNe).
   • Output: A continuous probability score (SN-score) representing the likelihood of an object being a supernova.
   • Performance: Achieved a ROC-AUC of 0.98 ± 0.01.

2. Active Anomaly Detection (PineForest):

   • Framework: PineForest (an implementation of Isolation Forest with expert feedback).
   • Innovation: The SN-score generated by the supervised classifier was added as an additional feature to the initial 47-feature set, creating an "augmented feature set."
   • Process:
     • An Isolation Forest is initialized.
     • An expert reviews the most anomalous objects, labeling them as Anomalous (A), Regular (R), or Unknown (U).
     • The model iteratively prunes trees inconsistent with expert feedback.
     • Priors: A small set of known SNe (as few as 10) can be provided as "priors" to immediately prune trees inconsistent with these examples, accelerating convergence.

3. Key Contributions

• Hybrid Feature Engineering: Successfully integrated a supervised probability score (SN-score) as a feature within an unsupervised anomaly detection framework. This allows the anomaly detector to autonomously determine optimal thresholds for SN-like behavior.
• Efficiency via Priors: Demonstrated that providing a minimal number of labeled priors (known SNe) significantly accelerates the discovery process in the PineForest framework.
• Scalable Discovery Strategy: Proposed a method that balances targeted discovery (finding SNe faster) with broad exploration (maintaining sensitivity to other rare anomalies like AGNs or variable stars).

4. Results

The methodology was tested across 10 ZTF fields, comparing four experimental setups:

1. Initial features without priors.
2. Initial features with priors.
3. Augmented features (with SN-score) without priors.
4. Augmented features with priors.

Key Findings:

• Statistical Significance: The combination of augmented features + priors yielded a highly significant improvement ($p \ll 0.05$) in SN recovery rates compared to other methods.
• Comparison with Pure Classification: The hybrid approach (Augmented + Priors) performed comparably to simply ranking objects by the SN-score alone (Spearman $\rho = 0.82$), but with the added benefit of retaining the ability to detect non-SN anomalies.
• Discovery Rate: Among the top 30 visually inspected candidates per field, the hybrid method with priors identified significantly more SNe than methods without priors or without the SN-score feature.

Specific Discoveries:
The application of this method to ZTF DR23 led to the identification of:

• 7 previously unreported SN candidates: Photometric modeling (SNCOSMO) classified them as 5 Type Ia, 1 Type IIn, and 1 Type IIP.
• 1 AGN candidate.
• 1 Unusual Galactic Variable (SNAD283): A helium-rich accretion system with a long-duration, low-amplitude outburst inconsistent with classical or dwarf novae.
• 2 Host Galaxies with Multiple Supernovae: Two galaxies hosting "supernova siblings" (multiple explosions in the same host), including a pair separated by 5 years (AT2018mpv and AT2023kuz) and another separated by 4 years (SN2018elp and AT2022mwi).

5. Significance and Future Impact

• Expert-Guided Efficiency: The study proves that incorporating domain knowledge (via SN-scores and priors) into unsupervised frameworks drastically reduces the "human-in-the-loop" effort required to find rare events in petabyte-scale datasets.
• Future Surveys: This approach is directly applicable to the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST), which will generate millions of alerts nightly. The hybrid method offers a scalable solution to prioritize scientifically interesting outliers without relying solely on rigid supervised classification.
• Balanced Discovery: Unlike pure classification, this method does not "blind" the system to other types of anomalies, ensuring that the search for SNe does not come at the cost of missing other novel astrophysical phenomena.