05 — Class imbalance

flowchart LR
    RARE["rare F/Q beats"] --> W["weight mistakes 50x\nforce into batches"] --> SEEN["model actually learns them"]

Class group	Rough share	What the model wants to do
N	the bulk	guess this always
V, S	small	learn if forced
F, Q	tiny (<1%)	ignore completely

The lazy trick (always guess common classes) scores well overall while catching nothing rare. Fixes used in heartbeat literature:

• Class weights: rare-class mistakes count extra (missing F can count 50x missing N).
• Weighted sampling: every training batch forced to include rare beats.
• Focal loss: auto-focuses learning on hard examples, down-weights easy ones.

Expectation: Q and F stay weak everywhere — all papers struggle there. Winning means N, V, S strong while F/Q degrade gracefully. That is honest science, and ../research/results.md will say so plainly.

Next: 06-ecg-signal-processing.md