Multi-feature Semi-supervised Learning Approach

Abstract

The concerns of this chapter are threefold: first, due to that each sensor feature can be exploited in different modality, multiple feature modalities for each sensor may be extracted; second, consider that the manual labeling of artificial olfaction data in real-time detection is difficult and hardly impossible, semi-supervised learning strategy is expected to be a breakthrough and overcome the problem of insufficient labeled data in artificial olfactory system; third, in E-Nose community, classifier learning is generally independent from feature extraction, such that the recognition capability of an E-Nose is limited due to the achieved suboptimal performance. Motivated by these concerns, in this chapter, from a new machine learning perspective, we aim at proposing a multi-feature kernel semi-supervised learning framework nominated as MFKS, whose merits can be composed of three points. (1) A multi-feature joint learning with low-rank constraint is developed for exploiting the multiple feature modalities from each sensor. The relatedness of all sub-classifiers learned on multiple feature modalities is preserved by imposing a low-rank constraint on the group classifier as regularization. (2) With a manifold assumption, a Laplacian graph manifold regularization is incorporated for semi-supervised learning and overcomes the flaw of insufficient labeled data in E-Nose. (3) The feature level and classifier level in artificial olfactory system are learned simultaneously in a complete framework, such that the recognition performance of an E-Nose can be optimally achieved. Experiments on two olfaction datasets including a large-scale 16-sensor data with 36-month drift and a small-scale temperature-modulated sensor data demonstrate that the proposed approach outperforms other algorithms.