Seonglae ChoMixed negative sampling
Negative Sampling for learning scores from large-scale data corpus. Dataset from the world is also unbalanced (positive interactions << negative interactions), so it tends to sample proper negative points rather than all negative pairs.
Batch Negatives (Unigram Sampling)
Uses other items in the same mini-batch as negative samples, which speeds up training but leads to popularity bias due to insufficient long-tail items.
2020 Google MNS (Unigram + Uniform Sampling)
Adding diverse negative samples improves recommendation quality.
Mixed Negative Sampling for Learning Two-tower Neural Networks in Recommendations
We maintain a portfolio of research projects, providing individuals and teams the freedom to emphasize specific types of work.
https://research.google/pubs/mixed-negative-sampling-for-learning-two-tower-neural-networks-in-recommendations/
