Comparison of NMF with Kullback-Leibler Divergence and Itakura-Saito Divergence for Odor Approximation

Odor approximation enables us to produce a variety odors by blending odor components. Approximated odor based on sensed odor is called odor reproduction. The reproduced odor should be as close to a target odor as possible. On the other hand, no primary odors have been found. Thus, finding an appropriate set of odor components with the minimum number to cover a wide range of odors is indispensable. Non-negative matrix factorization with Kullback-Leibler divergence (NMF-KL) and Itakura-Saito divergence (NMF-IS) were used to explore odor components using the mass spectrum data space of essential oils that we gathered using mass spectrometry (MS). The different property from both NMFs in treating small peaks were evaluated to explore odor components. The results shows that NMF-IS has a higher capability of odor reproduction compared with NMF-KL, especially in reproducing small peaks that appear in high m/z region.