Class imbalance and sampling

How this question breaks down

1. 1ImbalanceCore14 minWhen one example in 500 is a positive, accuracy, the 0.5 threshold and even your mini-batches stop telling you the truth; know which of them is actually broken before you resample anything.
2. 2ResamplingCore19 minDrop negatives, repeat or synthesise positives, or weight the loss. Each makes rare positives count for more, and each skews the model's probabilities, which you have to undo before the score is used as a probability.
3. 3NegativesCore20 minWith implicit feedback you never log a "no", so every negative is chosen; which items you choose (random, in-batch, popular or hard) decides what the model learns to tell apart, and each choice carries a bias you must correct.

Work through the parts in order and you’ve covered the full design. Each part is one topic.