This is companion repository to two tower models, a commonly used approach to retrieval / candidate generation in recommender systems. The goal of this repository is to show how to increase the alignment of retrieval with ranking.
This shows a sample implementation of two tower models in PyTorch.
This is text-book implementation of self attention and positional encodings to summarise user history for the user tower.
This uses the UserHistoryEncoder above to add to the user tower.
Since we are using net user value to estimate the relevance, this uses a weighting approach to position debiasing. There are other approaches as well like logit debiasing via shallow tower that we recommend trying out.
This extends TwoTowerWithPositionDebiasedWeights from above and implements a pointwise ranking module, broadly described in Revisting neural accelerators. We chose this since implementing a light ranking in the retrieval system is an intuitive way to increase consistency with main ranking.
Then we extend this to adding knowledge distillation from ranking models. This is similar to the approach described in How to reduce the cost of ranking by knowledge distillation
In this step we add a further layer of funnel consistency where, inspired by RLHF, we use the ranking model as a "reward model" and learn how to make the retrieval more aligned with the ranking model.
Disclaimer: These are the personal creations/opinions of the author(s). Any artifacts, opinions stated here are theirs and not representative of their current or prior employer(s). Apart from publicly available information, any other information here is not claimed to refer to any company including ones the author(s) may have worked in or been associated with.