Activity Number:
|
146
- Functional and High-Dimensional Analysis
|
Type:
|
Contributed
|
Date/Time:
|
Monday, August 8, 2022 : 10:30 AM to 12:20 PM
|
Sponsor:
|
ENAR
|
Abstract #323620
|
|
Title:
|
Deep Learning Aided Feature Selection for Highly Correlated Predictors: A Case Study with Diffusion-MRI Data
|
Author(s):
|
Arkaprabha Ganguli* and David Todem and Andrew Bender and Tapabrata Maiti
|
Companies:
|
Michigan State University and Michigan State University and Michigan State University and Michigan State University
|
Keywords:
|
Bootstrapping;
False Discovery Rate;
High Dimensional Variable Selection;
Lasso Penalty;
Deep Learning;
Penalized Regression
|
Abstract:
|
Machine learning and artificial intelligence profoundly benefit science and society as evidenced by their impressive success in many areas. This impact is fostered by massive datasets, ever-increasing computational resources, and the ability of deep neural network (DNN) to learn task-specific representations. The use of DNN models for feature selection has recently become a major area of research to take advantage of model complexity, while maintaining an explainable and parsimonious model. This paper considers the problem of feature selection with highly correlated high-dimensional predictors -- an established but commonplace source of challenges for traditional variable selection algorithms. We propose a novel screening and cleaning strategy with the aid of deep learning for cluster-level discovery of highly correlated predictors with controlled error rate. The extensive simulation study demonstrates the gain of the proposed method through its substantial enhancements to power while minimizing the number of false discoveries. Its application to an Alzheimer's disease tractography dataset produces anatomically meaningful discoveries in brain regions associated neurodegeneration.
|
Authors who are presenting talks have a * after their name.