A computer scientist passionate about building predictive systems and evaluating their impact. I enjoy dealing with large, messy data and have contributed to designing novel recommender systems and modeling their effect on information diffusion in social networks. As a Ph.D. student at Cornell, my research involves a combination of online experiments and large-scale data mining to deliver insights about people's preferences on items, and how recommender systems influence them. If you are interested, more details about my research are here: http://www.cs.cornell.edu/~asharma/Brief industry stints have helped me apply my research ideas to a larger scale. I developed a novel recommendation model for community recommendation at LinkedIn, prototyped collaborative models of augmenting restaurant data at Google and more recently, worked on estimating the impact of recommender systems in driving volume on e-commerce websites at Microsoft Research.
Causal principles for explaining machine learning: A session with DoWhy and DiCE libraries
What does it mean to explain a machine learning model? I will provide a definition of explanation based on causality that unifies different methods for model explanation. Methods like LIME and SHAP optimize for sufficient causes whereas counterfactual explanation methods illustrate the necessary causes of a model's output. The same definition also shows the pitfalls of current explanation methods and provides a way to develop explanations that are both faithful to the model and to the causal constraints that exist outside. I will showcase these ideas using two open-source libraries that provide causal explanations: the first (DoWhy) estimates the direct causal effect of changing any feature, and the second (DiCE) provides counterfactual explanations for any model output.