Odor Classification: Exploring Feature Performance and Imbalanced Data Learning Techniques

This research delves into olfaction, a sensory modality that remains complex and inadequately understood. We aim to fill in two gaps in recent studies that attempted to use machine learning and deep learning approaches to predict human smell perception. The first one is that molecules are usually represented with molecular fingerprints, mass spectra, and vibrational spectra; however, the influence of the selected representation method on predictive performance is inadequately documented in direct comparative studies. To fill this gap, we assembled a large novel dataset of 2606 molecules with three kinds of features: mass spectra (MS), vibrational spectra (VS) and molecular fingerprint features (FP). We evaluated their performance using four different multi-label classification models. The second objective is to address an inherent challenge in odor classification multi-label datasets (MLD)—the issue of class imbalance by random resampling techniques and an explainable, cost-sensitive multilayer perceptron model (CSMLP). Experimental results suggest significantly better performance of the molecular fingerprint-based features compared with mass and vibrational spectra with the micro-averaged F1 evaluation metric. The proposed resampling techniques and cost-sensitive model outperform the results of previous studies. We also report the predictive performance of multimodal features obtained by fusing the three mentioned features. This comprehensive and systematic study compares the predictive performance for odor classification of different features and utilises a multifaceted approach to deal with data imbalance. Our explainable model sheds further light on features and odour relations. The results hold the potential to guide the development of the electric nose and our dataset will be made publicly available.