Classifiers With a Reject Option for Early Time-Series Classification
This addresses the need for accurate early classification in signal processing, particularly for applications like electronic noses, though it appears incremental as it builds on existing ensemble methods with a reject option.
The paper tackles the problem of early time-series classification in dynamic environments by proposing a classifier with a reject option that makes online decisions without waiting for the full signal, using ensemble agreement instead of posterior probability. It applies this to odor classification, showing that the Forefront-Nose achieves robustness in earliness and recognition compared to standard classifiers in wind tunnel tests.
Early classification of time-series data in a dynamic environment is a challenging problem of great importance in signal processing. This paper proposes a classifier architecture with a reject option capable of online decision making without the need to wait for the entire time series signal to be present. The main idea is to classify an odor/gas signal with an acceptable accuracy as early as possible. Instead of using posterior probability of a classifier, the proposed method uses the "agreement" of an ensemble to decide whether to accept or reject the candidate label. The introduced algorithm is applied to the bio-chemistry problem of odor classification to build a novel Electronic-Nose called Forefront-Nose. Experimental results on wind tunnel test-bed facility confirms the robustness of the forefront-nose compared to the standard classifiers from both earliness and recognition perspectives.