Negative sampling
Choosing negative examples or sampled softmax with logQ correction for retrieval training.
45 cases from 27 companies apply Negative sampling. 22 were checked in an online A/B test; 8 report results in production. Includes variants: In-batch negatives, Sampled softmax.
Typical tasks: Retrieve and rank from millions of items, Recommend from what the user just did, Rank search results, Learn user and item embeddings and 8 more. Companies: Pinterest, Yandex, LinkedIn, Criteo, Airbnb, Expedia Group and 21 more. Often combined with: Two-tower model, In-batch negatives, Sampled softmax, Approximate nearest neighbors, Multi-task learning and 78 more.




