Comparing Different Dictionary-Based Classifiers for the Classification of Volatile Compounds Measured with an E-nose

Electronic noses (e-noses) are devices that mimic the biological sense of olfaction to recognize gaseous samples in a very fast and accurate manner, being applicable in a multitude of scenarios. E-noses are composed of an array of gas sensors, a signal acquisition unit and a pattern recognition unit including automatic classifiers based on machine learning. In a previous work, a text-based approach was developed to classify volatile organic compounds (VOCs) using as input signals from an in-house developed e-nose. This text-based algorithm was compared with a 1-nearest neighbor classifier with euclidean distance (1-NN ED). In this work we studied other text-based approaches that relied in the Bag of Words model and compared it with the previous approach that relied in the term frequency-inverse document frequency (TF-IDF) model and other traditional text-mining classifiers, namely the naive bayes and linear Support Vector Machines (SVM). The results show that the TF-IDF model is more robust overall when compared with the Bag of Words (BoW) model. An average F1-score of 0.84 and 0.70 was achieved for the TF-IDF model with a linear SVM for two distinct gas sensor formulations (5CB and 8CB, respectively), while an F1-score of 0.66 and 0.71 was achieved for the BoW model for the same formulations. The text-based approaches appeared to be less reliable than the traditional 1-NN ED method.