Wavelet-Based Deep Learning for Multi-Time Scale Affect Forecasting

Sy-Miin Chow, Young Won Cho, Xiaoyue Xiong, Yanling Li, Yuqi Shen, Jyotirmoy Nirupam Das, Linying Ji, Soundar R. Kumara

DOI: 10.64028/xnlk150705

Keywords: Longitudinal, machine learning, application

TL;DR: We present results using the scattering transform, a machine learning approach that integrates wavelet analysis with deep learning models to perform affect forecasting across multiple time resolutions.

Abstract: We present results using the scattering transform, a machine learning approach that integrates wavelet analysis with deep learning models in a single step, enabling efficient forecasting and classification. Because coefficients in the deep neural network are fixed to known coefficients in the wavelet analysis, computational burden and expenses are greatly reduced, with useful results found even with sample sizes that are comparably small for standard machine learning applications. Using illustrative and empirical examples designed to mirror multi-temporal and non-stationary changes in individuals' physiological and perceived (self-report) affect arousal, we propose a multi-subject extension of a feature activation heatmap proposed previously for convolutional network models, and illustrate its utility in displaying the time-varying importance of multiple physiological signals' frequency components in forecasting individuals' self-report affect arousal during a laboratory emotion induction task.

Submission Number: 11