A Novel Semi-Supervised Learning Approach in Artificial Olfaction for E-Nose Application

Abstract

Artificial olfaction data are usually represented by a sensor array embedded in an electronic nose system (E-Nose), such that each observation can be expressed as a feature vector for pattern recognition. The concerns of this paper are threefold: 1) each feature can be represented by multiple different modalities; 2) manual labeling of sensory data in real application is difficult and hardly impossible, which results in an issue of insufficient labeled data; and 3) classifier learning is generally independent of feature engineering, such that the recognition capability of E-Nose is restricted due to the unilateral suboptimum. Motivated by these concerns, in this paper, from a new perspective of multi-task learning, we aim at proposing a unified semi-supervised learning framework nominated as MFKS, and the merits are composed of three points. First, a multi-feature joint classifier learning with low-rank constraint is developed for exploiting the structural information of multiple feature modalities. The relatedness of sub-classifiers with respect to feature modalities is preserved by imposing a low-rank constraint on the group classifier. Second, with a manifold assumption, a Laplacian graph manifold regularization is incorporated for capturing the intrinsic geometry of unlabeled data. Third, the features and classifiers are learned simultaneously in a unified framework, such that the optimality and robustness are improved. Experiments on two data sets, including large-scale 16-sensor data with 36-month drift and small-scale temperature modulated sensory data, demonstrate that the proposed approach has 4% improvement in classification accuracy than others.