email:	dasta@stat.osu.edu
office:	427 Cockins Hall

I'm an associate professor at The Ohio State University in the Department of Statistics. Previously, I obtained a dual Ph.D. in Engineering and Public Policy (EPP) and Statistics under the supervision of my advisor, Cosma Shalizi at Carnegie Mellon University. Before that, I obtained an MPA from Institut d'études politiques de Paris (Sciences-Po) and a B.S. in Mathematics with a specialization in Economics from the University of Chicago. I'm interested in bringing geometric methods to bear on network inference, non-parametric non-Euclidean methods, and subsequent applications.

Network Inference

Network inference is about inferring a distribution of random graphs (networks) from sample graphs (networks). This is useful, for example, in assessing whether changes in networks over time, environment, or other conditions are significant or mere fluctuations. It is often natural to assume that the networks of interest are finite approximations of some interesting space. I am interested in the interplay between the geometry of the latent space (e.g. curvature) and the problems of network inference (e.g. estimator consistency).

Estimation on Curved Spaces

Data and their generative models are usually parametrized as collections of numbers, but both are often more naturally regarded as points in a non-Euclidean space. Some examples of such data are directional headings (the space SO(3) of 3x3 special orthogonal matrices), distortions in the spacetime continuum (symmetric positive definite matrices), or latent positions of large-scale networks (hyperbolic spaces). In all such examples, Euclidean distances do not reflect the true distance between the data, regarded as points in a true latent space. Similarly, parametric families of generative models are often studied as geometric spaces where the distance between two such generative models is their Fisher distance. I am interested in extending standard constructions and results regarding Euclidean statistics, such as kernel density estimation and maximum likelihood estimation, for the non-Euclidean setting.

Manifold Learning

Often, the space that the data is sampled from itself is unknown. For example, in computer graphics and machine vision, that space might literally be a geometric object one is trying to learn from imaging data. In another example, in social networks, that space might represent covariates for users with unknown correlations encoded by the hidden space. I am interested in developing tools for learning the geometry of the general space.