🔭 astrophysics

The Impact of Host Galaxy Properties on Supernova Classification with Hierarchical Labels

This paper demonstrates that host galaxy properties significantly enhance the photometric classification of supernovae, particularly when redshift information is unavailable, while also introducing a weighted hierarchical cross-entropy objective function and applying it to expand the Pan-STARRS Medium Deep Survey sample to over 4,400 events.

Original authors: V. Ashley Villar, Sebastian Gomez, Edo Berger, Alex Gagliano

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: V. Ashley Villar, Sebastian Gomez, Edo Berger, Alex Gagliano

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 the night sky is a massive, bustling city, and every night, new "fireworks" (supernovae) explode in the distance. For decades, astronomers have been able to identify these fireworks by sending a specialized team to the scene to take a close-up photo (spectroscopy). But with the upcoming arrival of a new, super-powerful telescope (the Vera C. Rubin Observatory), the number of fireworks will explode by a thousand times. The old team won't be able to keep up; they will only be able to visit a tiny fraction of the explosions.

So, astronomers need a new way to sort these fireworks instantly, using only the light they see from afar (photometry), without ever visiting the scene. This paper is about building a smarter sorting machine that uses a secret clue: the neighborhood where the firework happened.

Here is a simple breakdown of what the researchers found:

1. The Neighborhood Matters (Host Galaxies)

Think of a supernova like a person moving into a new house.

Type Ia Supernovae (the "thermonuclear" ones) are like retirees; they can live anywhere, in a mansion, a shack, or a suburb.
Core-Collapse Supernovae (the "massive star" ones) are like young, energetic artists; they only live in "artsy" neighborhoods where there is active construction and new life (star formation).

The researchers built a computer brain (a neural network) to look at the "neighborhood" (the host galaxy) to guess what kind of firework exploded.

The Result: If you just look at the neighborhood, the computer is incredibly good at spotting the "retirees" (Type Ia). It can find a group of them that is 90% pure (very few imposters).
The Catch: It's harder to tell the difference between the different types of "young artists" (like Type II vs. Type Ib/c) just by looking at the neighborhood.

2. The "Redshift" Map

Sometimes, astronomers know exactly how far away the firework is (redshift). This is like having a GPS coordinate.

With a GPS: If you know the distance and the neighborhood, the computer gets even better at spotting the "young artists" (Type II and Superluminous supernovae).
Without a GPS: If you don't know the distance, the neighborhood clues become super important. The computer relies heavily on the galaxy's color and size to make a good guess.

3. The "Light Curve" vs. The "Neighborhood"

The computer also looks at the "light curve"—the story of how the firework brightened and faded over time.

When we know the distance: The story of the light curve is so detailed that the neighborhood clues don't add much new information. It's like trying to guess a person's job by looking at their resume; once you read the resume, looking at their house doesn't tell you much more.
When we don't know the distance: The neighborhood clues become vital. Without the distance, the light curve story is a bit blurry, so knowing the "neighborhood" helps the computer fill in the gaps and make a much better guess.

4. A New Way to Grade the Test

The researchers also invented a new way to grade the computer's test scores.

Old Way: If the computer guessed "Type Ib" but the answer was "Type Ic," it got a zero. But both are "Type I" (no hydrogen), so it was a "close" mistake.
New Way (Weighted Hierarchical Cross-Entropy): This new grading system is smarter. It gives partial credit for "close" mistakes. If the computer guesses the right family of supernovae but the wrong subtype, it gets a better score. This encourages the computer to learn the family tree of supernovae rather than just memorizing a list.

5. The Big Update

Using these new methods, the team went back and re-sorted a massive list of 4,400+ fireworks from a previous survey (Pan-STARRS) that didn't have distance information.

They found that their new method agrees well with old methods for the common types (Type Ia and Type II).
However, for the rare and exotic fireworks (Superluminous supernovae), the new method and the old methods often disagreed. This suggests that without knowing the distance, it is very hard to reliably identify these rare, bright events.

The Bottom Line

In the future, when we have too many supernovae to count, we can't rely on just the light curve. We need to look at the host galaxy (the neighborhood).

If we know the distance, the neighborhood helps a little bit.
If we don't know the distance, the neighborhood is a critical clue that makes our classification much more accurate.

The paper concludes that while we can't perfectly identify every rare type of supernova without a spectroscopic visit, we can use the "neighborhood" to create very pure, reliable lists of the most common types, which is essential for studying the universe on a massive scale.

Technical Summary: The Impact of Host Galaxy Properties on Supernova Classification with Hierarchical Labels

Problem Statement
With the impending advent of the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST), the discovery rate of supernovae (SNe) is projected to exceed the rate of spectroscopic follow-up by three orders of magnitude. Currently, only $\sim10\%$ of SNe receive spectroscopic classification, a figure expected to drop to $\lesssim0.1\%$ without additional resources. This disparity necessitates robust photometric classifiers capable of selecting events for follow-up in real-time and enabling archival population studies. While photometric classification based on light curves is established, the utility of contextual host galaxy information remains under-explored, particularly regarding its ability to improve classification accuracy in the absence of redshift data and its interaction with hierarchical taxonomies.

Methodology
The authors utilize a sample of 557 spectroscopically identified SNe from the Pan-STARRS Medium Deep Survey (PS1-MDS), comprising five classes: Type Ia, Type II, Type IIn, Type Ib/c, and Type I superluminous SNe (SLSNe). This sample serves as a realistic analogue for LSST data in terms of cadence and depth.

Feature Sets: The study evaluates six distinct classifier configurations using a Multi-Layer Perceptron (MLP):
1. Host galaxy features only (without redshift).
2. Host galaxy features + redshift.
3. Light curve features only (without redshift).
4. Light curve features + redshift.
5. Host galaxy + light curve features (without redshift).
6. Host galaxy + light curve features + redshift.
Host galaxy features are extracted via the FLEET pipeline and include magnitudes, radii, separation offsets, and probability scores derived from Pan-STARRS 3 $\pi$ survey data. Light curve features are drawn from SuperRAENN, including non-linear autoencoder features, rise/decline times, and peak magnitudes.
Classification Architecture: The authors introduce a novel objective function, the Weighted Hierarchical Cross-Entropy (WHXE), to address the hierarchical nature of SN taxonomy (e.g., distinguishing Core-Collapse from Thermonuclear before subclassifying). The graph structure includes a root node ("Supernova"), branching into Type Ia and Core-Collapse, with further subdivisions for H-rich (Type II) and H-poor (Type Ib/c, SLSN) classes. The WHXE loss function penalizes misclassifications based on their distance in the hierarchy, encouraging the model to place objects in the correct broader category even if the specific subclass is incorrect.
Training: The MLP is trained using the Adam optimizer with hyperparameter optimization (learning rate, batch size, hierarchy weighting $\alpha$ , and neuron count). Missing data is handled via K-means imputation.

Key Results
The performance of the classifiers is evaluated using purity, completeness, and F1-scores across five-way, three-way, and binary classification tasks.

Host Galaxy Information Alone:
- Type Ia SNe: Host galaxy properties alone can isolate relatively pure ( $>90\%$ ) samples of Type Ia SNe, even without redshift information.
- Core-Collapse SNe: With redshift information, host properties allow for the production of somewhat pure ( $>70\%$ ) samples of Type II SNe and SLSNe.
- Limitations: Without redshift, the classification of Type IIn and Type Ib/c SNe using only host properties remains poor (purity $\lesssim0.2$ ).
Impact of Redshift:
- Without Redshift: Host galaxy properties significantly increase the accuracy of photometric classification across all subtypes. In the absence of redshift, galaxy features are among the top 10 most important features in a Random Forest analysis.
- With Redshift: When complete light curves and redshift information are available, adding host galaxy properties does not significantly improve classification accuracy for most subtypes. The only statistically significant improvement is a slight increase in the purity of Type Ib/c SNe, though this comes with a marginal decrease in completeness.
Hierarchical Classification:
- The WHXE objective function successfully trains a multi-layered classifier. In binary classification (Type Ia vs. Core-Collapse), the host-only classifier achieves $\sim74\%$ accuracy with redshift, matching state-of-the-art results.
- In three-way classification (Type Ia, H-rich, Stripped-Envelope), the host-only classifier achieves an F1-score of 0.54 with redshift, outperforming previous literature that relied on early light curves.
Application to PS1-MDS:
- The authors apply their redshift-independent classifier to 4,407 SN-like transients from the PS1-MDS lacking spectroscopic redshift.
- Discrepancies: The new classifier shows strong agreement ( $>80\%$ ) with previous classifiers (SuperRAENN, SuperPhot) for Type Ia and Type II SNe. However, agreement is significantly lower for SLSNe and Type IIn SNe. Notably, many objects previously classified as SLSNe are reclassified as Type Ia or AGN by the new model, suggesting that without redshift, photometric classifiers struggle to produce consistent, pure SLSN samples.

Significance and Claims
The paper claims to provide the first formal application of the Weighted Hierarchical Cross-Entropy loss function to supernova classification, offering a more natural accounting of the hierarchical taxonomy of transients.

The primary scientific contribution is the quantification of host galaxy utility:

Host galaxy information is a powerful tool for isolating Type Ia SNe and distinguishing Core-Collapse from Thermonuclear events, even in the absence of redshift.
However, for detailed subtyping (e.g., distinguishing Ib/c from II or identifying SLSNe) when full light curves and redshifts are available, host galaxy properties offer diminishing returns.
The study highlights the critical need for accurate photometric redshifts and robust host association techniques (like GHOST) to maximize classification efficiency in the LSST era.
Finally, the work serves as a cautionary note for population-level studies: agreement between different photometric classifiers is highly variable, particularly for rare classes like SLSNe, necessitating careful calibration of classification confidence and bias correction.