Two-stage RecSys at Unity Ads – Denis Vorotyntsev, Unity Technologies


Session Outline

A two-stage system produces recommendations in two steps: a nominator presents a small number of candidates selected from a larger pool. Then, the ranker rearranges nominated items and serves them to the user. In this talk, Denis Vorotyntsev shows the motivation for having a two-stage system at Unity. Denis presents the overview of the pipeline and architecture of the ML part. He also shares tricks they used to improve the quality of the solution and pitfalls they have faced.

Key Takeaways

  • Saving GPU resources or When and why do we serve fancy DL on CPU
  • Measuring the goodness of a two-stage system or Why logloss is not all you need
  • Making results significant again, or How do we run AB tests

Add comment