Periodically variable star classification from light curves.
Variable star classification is a real signal-processing problem dressed up as ML. The notebook treats it as a feature-engineering exercise on the light curves, then a classical classifier — every step visible so someone else can reproduce the result.
I extracted phase-folded amplitude, periodicity strength, and shape descriptors from the raw light curves. Each feature has a notebook cell that plots what it looks like for the major classes — so the model isn't a black box, it's reading the same signals an astronomer would.
A scikit-learn pipeline does the actual fit — deliberately classical, so the result can be inspected, retrained on new data, and explained. The notebook ships with the train/val split, the metrics that matter for an imbalanced star catalogue, and the confusion matrix that shows where the model still struggles.
End-to-end notebook: ingest light-curve catalogue → engineered features → train classifier → confusion-matrix readout.
Classical features over a black-box CNN — the signal IS a periodogram. Stratified split because the class balance is skewed.
Hand-crafted features take longer to engineer. Classical model has a hard ceiling on rare-class recall.
0.92 accuracy across four star classes, 0.89 macro-F1, and every step reproducible from the notebook in one run.
The public GitHub repository is the canonical home for Applied research. You can read the commit log, fork it, file an issue, or send a PR.