[Baselines] Add Baseline ML Models For Time Domain Data

[benboyd97]

Add Baseline ML Models For Time Domain Data

Add baseline ML models and metrics to evaluate time domain tasks using #17 data.

Contacts: Participants: Helen Qu

Goals and deliverable

[describe your goals for the week and the deliverables you are aiming for]

Resources needed

[describe the resources (software, skills, data, or just enthusiasm) needed for this project] Data:

• TESS
• Kepler

Detailed description

Possible tasks:

• Photometric object classification
  • Avocado, Connect Later, SCONE, RAPID/SCONE (early-time SN classification)
• Spectroscopic classification
  • DASH
• Anomaly detection on datasets that are already tagged?
• Regression of missing lightcurve data
• Connections between host and transient type
• Host matching with transients?
  • GHOST, DELIGHT, DLR (non-ML)
• Stellar variability classification (e.g., starclass, see Audenaert et al. (2021))
• parameter inference (e.g., stellar rotation rate, photo-z, supernova stretch/color)

Metrics:

• Weighted multi-class logarithmic loss for SNe classification
• Mean squared error for missing continuous data - like missing lightcurve points
• Uncertainty quality (calibrating bayesian approximations?)

[EiffL]

We were talking with @jeraud about tasks on stellar light curves and looking specifically at this paper: https://arxiv.org/abs/2309.16316 And here is the code: https://github.com/panjiashu/Astroconformer

Here is an example of regression of log g:

[jeraud]

• https://arxiv.org/abs/2308.06404
• https://www.aanda.org/articles/aa/full_html/2023/02/aa43928-22/aa43928-22.html

[David-Chemaly]

Added XGBoost baseline for SN classification

[David-Chemaly]

We finished developing a straight forward baseline XGBoost model for classification of SN types and prediction of photo-redshift. The baseline works and has been pushed to the YSE branch!!

We also tried a Transformer approach but that sadly proved to be a bit more demanding for a baseline.