Improvement of Odor Impression Predictive model using Machine Learning

In the Sensory test to observe human impression for an odorant molecule, it is difficult to obtain reliable data because of its cost and complicated structure of odor perception space. However, in the previous studies, we proposed a model to predict odor impression from mass spectrum using proposed DNN. However, the accuracy of our model was still insufficient and further improvement was needed. In this study, we've studied two methods of using a large-scale dataset for training auto encoder for mass spectrum and Itakura-Saito divergence as a cost function. As a result, the correlation coefficient between predicted and true values was raised from 0.76 to 0.90.