Back to Projects

Build from first principles

Build a Two-Tower Recommender

Learn user and item vectors with two towers, retrieve candidates fast, and stop popular items from crowding out everything else.

What is a two-tower recommender?Two networks, one dot product, and why retrieval comes before ranking

1Score Users Against ItemsScore every user against every item

2Train With In-Batch NegativesLearn from the other items in the batch

3Correct In-Batch Sampling Bias With log qStop punishing popular items for being popular

4Retrieve the Top K Unseen ItemsPick candidates a user has not seen

5Measure Recall@K, Tail Recall and CoverageJudge the retrieved lists

From Towers to RecommendationsSee how your pieces train and judge a real retrieval model

Train, Break and Fix a RetrieverTrain two towers, watch in-batch negatives hide popular items, correct with log q, and weigh recall against coverage