A Semi-Supervised Learning Approach to Artificial Olfaction

Abstract

In the last decade, semi-supervised learning (SSL) has gained an increasing attention in machine learning. SSL may obtain performance gains by adding to the supervised information, provided by a limited labelled training set, the information content embedded in an unsupervised sample set. This may be very helpful, since obtaining supervised samples can be difficult and costly, as in several artificial olfaction (AO) problems. In this work, co-training style semi-supervised algorithms are applied to air pollution monitoring, an on-field artificial olfaction problem. The primary purpose is to adapt a regressor knowledge to the well known sensors and concept drift issues that characterize the use of solid state chemical sensors in harsh environments. The response of an array of solid state chemical sensors, located in a city street affected by heavy cars traffic, has been monitored for more than 1year and used to estimate hourly pollutants concentrations. Conventional analyzers provided the needed ground truth. Results obtained by the proposed approach show that it can both reduce the number of labeled samples needed for the multivariate calibration of the device and the performance decay due to drift effects.