The Fifth Elephant 2017

The art of trading between exploiting the best arm versus exploring for further knowledge of other arms has long been studied as Bandit Algorithms in various fields of clinical trials, designing financial portfolios, etc. Recently, in website optimization, these algorithms have been used for optimizing click through rates and performing A/B testing. However, these algorithms has the potential to be applied at several other contexts where we need to optimize for reward by exploring possible arms.

In this talk, I would be presenting an approach based on Contextual Multi-armed Bandit algorithms for achieving a better tradeoff between user’s expectation of faster replies and doctor’s burnout rate on a QnA platform like Practo Consult.

Background:

Practo’s Consult platform helps users get their health queries clarified by professional qualified doctors. As we value both users and doctors as our customers, we optimize for enhancing their experience while using Consult. This would entail good quality and faster replies for user’s health queries, whereas doctors expect good quality questions and number of assignments to be correlating with their answering capacity. This is an interesting problem because optimizing for one would compromise the other. For example, if we assign all questions to only the high performing doctors, the remaining non-performing doctors cannot undo their behavior as they won’t get enough questions. Whereas, if we balance the assignment of questions among doctors, it won’t be optimizing for faster replies. Ideally, we want to achieve a tradeoff between the two.

Outline

1. The dynamics of a QnA platform like Practo Consult
2. Introducing Multi-armed Bandit algorithm
3. Adapting a version of Bandit algorithm called Contextual Multi-armed Bandit to enhance the experience of users and doctors.

Speaker bio

Santosh GSK is working as a Senior Data Scientist at Practo. He has 5 years of industry experience in Data Science and 3 years as a ML Researcher with half a dozen publications in leading conferences. He is currently working on building data-driven solutions to improve both patient and doctor experience at Practo. Prior to that, he was working as a Data Scientist at Housing.com, where he worked on lead prediction and property price prediction models.

Links

• https://www.linkedin.com/in/santosh-gsk-a6412b2a/

Slides

https://drive.google.com/open?id=0B_FveDU9pdasdjhpVDZpMXRKNFk