A Perception-Driven Framework for Predicting Missing Odor Perceptual Ratings and an Exploration of Odor Perceptual Space

Abstract

There have been abundant research efforts on predicting verbal descriptions of odorants through the physicochemical features, physiological signals and E-nose signals. These approaches are interpreted as feature-driven methods in which the information about the inner links among different odor percepts is ignored. Different from that, we propose a perception-driven framework for predicting the missing odor perceptual ratings from other known odor percepts. Specifically, the work emphasizes pleasantness prediction based on level of importance in odor perception. In essence, this approach utilizes the relations among different odor perceptions, exploring the odor perceptual space subsequently. The missing perceptual ratings are predicted with an accuracy higher than 0.5 for more than half of the odor verbal descriptors, and almost half of the descriptors are predicted with a correlation higher than 0.8. The asymmetric clustering structure of odor perceptual space is revealed by feature selection for predicting the missing perceptual ratings. It is found that `pleasantness' is primarily determined by `sweet'.

Keywords

• Feature selection
• perception-driven
• perceptual ratings
• pleasantness