Notes

Learning-to-rank features that actually mattered

Most of the gain came not from model architecture but from a handful of well-normalised features: freshness, query-item text match, and per-query relative price. Feature leakage from future clicks was the most common bug.