Amit Sharma

À propos

Data tells stories. My research aims to tell the causal story.

I am a researcher at Microsoft Research India. From 2015-2017, I was a postdoctoral researcher at MSR New York working with Jake Hofman et Duncan Watts. Prior to that, I completed my Ph.D. in computer science at Cornell University (fortunate to be advised by Dan Cosley) and my undergraduate at IIT Kharagpur. My work has received a Best Paper Honorable Mention Award at the 2016 ACM Conference on Computer Supported Cooperative Work and Social Computing (CSCW), the 2012 Yahoo! Key Scientific Challenges Award and the 2009 Honda Young Engineer and Scientist Award.

The key insight in my work is to consider modern algorithms as interventions, just like a medical treatment or an economic policy. As machine learning systems move into societally critical domains such as healthcare, education, finance and criminal justice, questions on their impact gain fundamental importance. Unlike typical interventions studied in social and biomedical sciences, however, algorithmic interventions can be arbitrarily complex. I work on developing methods to estimate causal impact of such systems and build algorithms that optimize the causal effect. I am also passionate about designing new interventions for societal impact, especially in low-resource healthcare.

Take recommender systems as an example. Algorithmic recommendations are ubiquitous in the online world, celebrated for providing roughly 30-80% of activity and yet criticized for restricting people’s information access in social media to “filter bubbles”. A natural question then is: How much do these systems actually cause this behavior, and how much of it is just correlational? My research shows that these claims overstate the causal effect: e-commerce recommendation engines such as Amazon’s account for less than 10% of the traffic, and recommendation systems that show friends’ activities in music, movies, or books, account for a tiny fraction of people’s overall activity in these domains.

Two central questions in my research are:

• How do we correctly design algorithmic interventions, reason about generalizability, and plan for when they go wrong?
• How do we design a new algorithmic intervention to have the desired effect?

Most of my current work is on the first question, where I build on causal inference literature and develop methods that require minimal assumptions to estimate a causal effect. The second one is harder, and I look forward to a future where we can predict causal effect of an algorithmic intervention (e.g., routing supporters on a mental health support forum). While there are ongoing efforts in machine learning and statistics to optimize algorithmic interventions through better evaluation criteria or data collection, a purely algorithmic approach just scratches the surface of what is really a sociological problem—the role of algorithms embedded with human decision-making in a societal context. Put differently, the fundamental problem remains a causal one with interacting effects of between societal context, algorithms and individuals.

At Microsoft Research, I work on these problems with an amazing set of colleagues from the Machine Learning and AI and the Technology and Empowerment groups.