Authors: Yuta Saito, Himan Abdollahpouri, Jesse Anderton, Ben Carterette, Mounia Lalmas

Published on: April 24, 2024

Impact Score: 7.4

Arxiv code: Arxiv:2404.15691

Summary

• What is new: A new framework called Long-term Off-Policy Evaluation (LOPE) that estimates long-term outcomes of an algorithm using historical and short-term experiment data.
• Why this is important: Difficulty in estimating the long-term outcomes of algorithms due to delayed effects, which makes algorithm selection slow.
• What the research proposes: LOPE, which uses reward function decomposition to leverage short-term rewards for estimating long-term outcomes under a relaxed set of assumptions.
• Results: LOPE outperforms existing methods, especially when existing assumptions do not hold and in the presence of noisy long-term rewards. It proved more accurate in real-world tests on a music streaming platform.

Technical Details

Technological frameworks used: Long-term Off-Policy Evaluation (LOPE)

Models used: nan

Data used: Historical and short-term experimental data, real-world A/B test data from a music streaming platform

Potential Impact

Online platforms and companies relying on user engagement metrics could benefit, including music streaming services, e-commerce, and social media platforms.

Want to implement this idea in a business?

We have generated a startup concept here: FutureSight Analytics.