As a part of the Center for Science of Information Spring 2020 Seminar Series (online), Mohsen Heidari, Postdoctoral Fellow, Center for Science of Information, will be presenting a seminar on "Feature Selection for Supervised Binary Classification." ​

A Fourier-Based Framework for Feature Selection and Supervised Binary Classification

Abstract:
In learning datasets with a large number of attributes, feature selection has been an effective solution to reduce running time, improve learning accuracy, and facilitate a better understanding of the learning model. In feature selection, formulating a computationally efficient and theoretically justified measure to evaluate the feature subsets is a challenging task.  In theory, provable relations between the conventional feature selection measures and the prediction accuracy remain open. We aim to address this challenge by taking an information-theoretic perspective. For that, we develop a Fourier-based framework to study the feature selection problem. In this talk, I present this framework and describe our ongoing efforts to this end.  Specifically, we propose a Fourier-based learning algorithm with an embedded feature selection for supervised binary classification.   In binary classification problem involving d features, given a parameter ­ k < d , our algorithm selects k features and generates a predictor which depends only on the selected features. We propose a class of Fourier-based measures for feature selection and drive their connection to classification accuracy. As for theoretical guarantees, we derive a set of conditions for the optimality of the algorithm.

Recorded seminar video: https://www.youtube.com/watch?v=nvoWi3DXOM0