Conservative Policy Construction Using Variational Autoencoders for Logged Data With Missing Values.

Area of Science:

• Decision-making under uncertainty
• Machine learning for healthcare
• Data imputation techniques

Background:

• High-stakes decision-making requires policies that maximize rewards while minimizing risks.
• Online exploration is often infeasible in critical applications like healthcare, necessitating the use of observational data.
• Observational datasets frequently contain missing values, complicating policy learning.

Purpose of the Study:

• To develop a method for constructing personalized policies using logged data with missing feature attributes.
• To recommend optimal actions (treatments) even when observing a degraded version of features with missing values.
• To introduce and evaluate strategies for handling missing data in policy learning.

Main Methods:

• Investigated three strategies for addressing missing data in policy construction.
• Introduced a 'conservative strategy' designed to manage uncertainty arising from missingness.
• Utilized partial variational autoencoders (PVAEs) to estimate the posterior distribution p(X| ~ X) and capture data structure.

Main Results:

• The proposed conservative strategy effectively handles uncertainty caused by missing feature attributes.
• PVAEs demonstrated capability in modeling the underlying structure of features with missing values.
• The developed methods enable personalized policy construction from imperfect observational datasets.

Conclusions:

• Personalized policies can be learned from observational data with missing values using robust strategies.
• The conservative approach, implemented with PVAEs, offers a safe method for decision-making under data uncertainty.
• This work advances the application of data-driven decision-making in critical domains like healthcare.