Reproduce “Off-Policy Evaluation and Learning for Matching Markets” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yudai Hayashi, Shuhei Goda, Yuta Saito
New off-policy evaluation and learning methods, DiPS and DPR, leverage the two-stage reward structure inherent in matching markets to provide more accurate and reliable performance estimates than traditional methods. These approaches demonstrate significantly lower MSE and variance across synthetic and real-world job-matching datasets, enabling more efficient policy development at Wantedly, inc.
Key insights
- Modeling matching market interactions as a two-stage reward process (initial engagement, then mutual response) is crucial for effective off-policy evaluation.
- Applying importance weighting to the more frequent first-stage reward, rather than the extremely sparse ultimate match reward, significantly reduces variance in OPE.
- Combining direct method (regression) for the second-stage reward with importance weighting for the first-stage allows for robust estimators that balance bias and variance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.