On the 14th of August 2017 Prof. Tsamardinos will give an invited talk at the 2017 ACM SIGKDD Workshop on Causal Discovery. Title: “Advances in Causal-Based Feature Selection”
The talk will illustrate the past and explore the future of feature selection algorithms based on Causal Discovery theory. These algorithms allow the identification of the predictors, among thousands of potential candidates, that influence the most a given phenomenon, for example patients’ survival or customers’ churn.
SIGKDD is one of the major conference in Machine Learning and Data Analysis, placing a strong emphasis on business and real world applications.
Feature selection (a.k.a. variable selection) is a common task in data analytics, where the goal is to identify a minimal-size, optimally predictive feature subset. Theoretical results connect the solutions of the problem with the causal mechanism that generated the data, often represented by a Bayesian Network, a Maximal Ancestral Graph, or a Semi-Markov Causal Network. In such frameworks, the selected features are not only predictive of a target outcome of interest but also have a causal interpretation. Such results have given rise to a class of algorithms inspired by causal modeling of the data distribution. In the talk, we examine prototypical causally-inspired feature selection algorithms, advances that allow the algorithms to scale to high-dimensional problems, be applicable to a plethora of different types of data, identify multiple statistically-equivalent solutions, and scale to Big Data.