Method

Sampled softmax

Softmax over sampled negatives with logQ correction.

10 cases from 5 companies apply Sampled softmax. 5 were checked in an online A/B test; 1 report results in production.

Typical tasks: Retrieve and rank from millions of items, Recommend from what the user just did, Reduce ML cost, Balance relevance and diversity and 1 more. Companies: Yandex, Pinterest, Netflix, Expedia Group and 1 more. Often combined with: Two-tower model, In-batch negatives, Approximate nearest neighbors, SASRec, Context engineering and 27 more.