Avatar

Sanket Shah

PhD Student

Harvard University

⚠️ I am looking for postdoctoral positions! ⚠️

I am a fifth-year PhD student at Harvard University advised by Prof. Milind Tambe. I am broadly interested in decision-making under uncertainty by using techniques like Reinforcement Learning, Differentiable Optimization, and Task-Specific Loss Functions.

My PhD research focuses on Decision-Focused Learning—training and evaluating models by the quality of their downstream decisions, as opposed to their predictive accuracy—and its applications to public health. In collaboration with ARMMAN, an Indian NGO that uses mobile health programs to improve maternal and neonatal health, I’ve been part of a team that helped develop and evaluate an intervention planning system. This technology has been deployed in the field and has positively impacted the lives of more than 350,000 women.

Previously, I was Research Engineer at Singapore Management University (SMU) advised by Prof. Pradeep Varakantham where I used Reinforcement Learning to solve problems in Transportation and Security. I also spent a year at Microsoft Research India during which I worked on Information and Communications Technology for Development (ICTD) with Dr. Colin Scott and Dr. Bill Thies, and Natural Language Processing with Dr. Sundararajan Sellamanickam.

Interests

  • Algorithmic Decision-Making
  • AI for Social Good
  • Machine Learning
  • Reinforcement Learning

Education

  • Ph.D. in Computer Science, 2020 - Present

    Harvard University

  • B.E. (Hons.) in Computer Science, 2017

    Birla Institute of Technology and Science, Pilani

Recent Publications

Evaluating the Effectiveness of Index-Based Treatment Allocation

We show how a connection to the ‘individualized treatment rules’ literature in statistics to create asymptotically valid confidence intervals for RCTs that measure the quality of ‘clever’ intervention policies.

Group Fairness in Predict-Then-Optimize Settings for Restless Bandits

We show how decision-focused learning can improve the fairness-accuracy tradeoff in algorithmic decision-making using a case study on a real-world domain.

Efficient Public Health Intervention Planning Using Decomposition-Based Decision-Focused Learning

We propose an efficient way to implement decision-focused learning for the kinds of RMABs used for intervention planning in public health.

Leaving the Nest: Going Beyond Local Loss Functions for Predict-Then-Optimize

We propose two new innovations to help improve the learning of task-specific loss functions.

Preliminary Results in Low-Listenership Prediction in One of theLargest Mobile Health Programs in theWorld

We perform preliminary analysis on the Kilkari mobile health program and find that past approaches to predicting listenership in mobile health programs don’t work well.

Recent Posts

Learning Loss Functions for Predict-then-Optimize

A blog post about the predict-then-optimize paradigm for algorithmic decision-making, and how we can view it as inducing a …

My Ride-pooling Journey

A blog post about the joyride my paper ‘Neural Approximate Dynamic Programming for On-Demand Ride-Pooling’ took me on.