In-Batch Negatives and Popularity Bias in Two-Tower Training

You train a two-tower retrieval model with in-batch softmax negatives, batch size 1024. After training, you notice the model retrieves extremely popular items for almost every user, even users whose history suggests niche interests.

1. Explain the mechanism by which in-batch negative training biases the model toward popular items.
2. Describe the logQ correction and how it fixes this.
3. Would adding hard negatives (items a previous production model retrieved but the user didn't click) fix the same problem, a different problem, or both? Explain.