propensity score estimate

[arita37]

Hello,

In the input dataset, propensity scores needs to be provided, does the the propensity score needs to be calibrated ?

How is the impact of wrong propensity score ? (vs reward level...)

[usaito]

@arita37 Calibrating your pscore estimate might help, but depends on the situation. The following papers explore the effect of calibration in OPE, so might be of your interest:

• Aniruddh Raghu, Omer Gottesman, Yao Liu, Matthieu Komorowski, Aldo Faisal, Finale Doshi-Velez, Emma Brunskill. Behaviour Policy Estimation in Off-Policy Policy Evaluation: Calibration Matters. https://arxiv.org/abs/1807.01066
• Yuta Saito, Takuma Udagawa, Haruka Kiyohara, Kazuki Mogi, Yusuke Narita, Kei Tateno. Evaluating the Robustness of Off-Policy Evaluation. https://arxiv.org/abs/2108.13703.

If you're working on a particular application, you might want to evaluate the effect of calibrated vs non-calibrated pscore estimate on the OPE accuracy using synthetic data that mimic your real data. Should be easily implementable with OBP.

[arita37]

Ok.

But, if you do the math, no calibrated proba can create significant bias.

Different calibration strategies can be used with different temperatures

On Sep 27, 2022, at 3:58, yuta-saito @.***> wrote:

 @arita37 Calibrating your pscore estimate might help, but depends on the situation. The following papers explore the effect of calibration in OPE, so might be of your interest:

Aniruddh Raghu, Omer Gottesman, Yao Liu, Matthieu Komorowski, Aldo Faisal, Finale Doshi-Velez, Emma Brunskill. Behaviour Policy Estimation in Off-Policy Policy Evaluation: Calibration Matters. https://arxiv.org/abs/1807.01066 Yuta Saito, Takuma Udagawa, Haruka Kiyohara, Kazuki Mogi, Yusuke Narita, Kei Tateno. Evaluating the Robustness of Off-Policy Evaluation. https://arxiv.org/abs/2108.13703. If you're working on a particular application, you might want to evaluate the effect of calibrated vs non-calibrated pscore estimate on the OPE accuracy using synthetic data that mimic your real data. Should be easily implementable with OBP.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.