Predicting individual perceptual scent impression from imbalanced dataset using mass spectrum of odorant molecules

Predicting odor impression is considered an important step towards measuring the quality of scent in the food, perfume, and cosmetic industries. In odor impression identification and classification, the main target is to predict scent impression while identifying non-target odor impressions are less significant. However, the effectiveness of predictive models depends on the quality of data distribution. Since it is difficult to collect large scale sensory data to create an evenly distributed positive (target odor) and negative (non-target odor) samples, a method is necessary to predict the individual characteristics of scent according to the number of positive samples. Moreover, it is required to predict large number of individual odor impressions from such kind of imbalanced dataset. In this study, we used mass spectrum of flavor molecules and their corresponding odor impressions which have a very disproportioned ratio of positive and negative samples. Thus, we used One-class Classification Support Vector Machine (OCSVM) and Cost-Sensitive MLP (CSMLP) to precisely classify target scent impression. Our experimental results show satisfactory performance in terms of AUC(ROC) to detect the olfactory impressions of 89 odor descriptors from the mass spectra of flavor molecules.