TF-Ranking: Scalable TensorFlow Library for Learning-to-Rank

Topics: AI (Deep Learning), Learning-to-rank, Marc Najork, Navboost, Ranking, Reranking, User Signals

This research paper by Google presents TensorFlow Ranking (TF-Ranking), a scalable open-source library designed by Google to address ranking problems using deep learning. Ranking, a problem in information retrieval, focuses on optimizing the order of results (such as search engine results) rather than predicting labels or values like in classification or regression. The library provides flexible APIs to implement various scoring mechanisms, loss functions, and evaluation metrics to enhance ranking models. The library is optimized for large-scale applications and integrates features like unbiased learning-to-rank, using Inverse Propensity Weights (IPW) to correct bias in click data. TF-Ranking has been empirically tested on large datasets for Google services like Gmail search and Google Drive recommendations, showing improvements in ranking performance by using listwise losses over pointwise or pairwise approaches.

Participants: Rama Kumar Pasumarthi, Xuanhui Wang, Cheng Li, Sebastian Bruch, Michael Bendersky, Marc Najork, Jan Pfeifer, Nadav Golbandi, Rohan Anil, Stephan Wolf

Company: Google