CRII: RI: Towards a Comprehensive Dynamic Subset Selection Framework

Northeastern University, Boston MA

Investigators

Abstract

Subset selection is the task of finding a small subset of informative items from a large ground set. Sequential data, including time-series and ordered data, form an important large part of modern datasets, requiring effective subset selection techniques. This project develops a unified framework for sequential subset selection that incorporates dynamic models and relationships among items, addresses both unsupervised and supervised problems and handles multi-modal data. The project develops practical tools for video and multi-media summarization, human information selection models that benefit human-in-the-loop systems, as well as providing interdisciplinary research training to students and disseminates research results. The research results from this project can impact several communities, such as machine learning, computer vision, signal processing, visualization, and health-informatics. This research develops a comprehensive framework for subset selection in sequential datasets by addressing three important problems: i)it develops novel algorithms for sequential subset selection that incorporate dynamic models and relationships among sequential data and develops efficient optimization techniques to solve the problem; ii) it addresses the problem of unsupervised and supervised sequential subset selection by generalizing metric learning for subset selection and by tackling the challenge of the lack of pairs of positive and negative items; and iii) it proposes a multi-modal sequential subset selection framework to effectively take advantage of all modalities in datasets. The project brings together tools from sparse and low-rank recovery, convex optimization, message passing, metric learning and dynamical systems to tackle the problem. The project also designs objective evaluation methods to measure the performance of the methods.