Stream-based active learning with linear models

The proliferation of automated data collection schemes and the advances in sensorics are increasing amount we able to monitor real-time. However, given high annotation costs time required by quality inspections, is often available an unlabeled form. This fostering use active learning for development soft sensors predictive models. In production, instead performing random inspections obtain product information, labels collected evaluating information content data. Several query strategy frameworks regression have been proposed literature but most focus has dedicated static pool-based scenario. this work, propose a new stream-based scenario, where instances sequentially offered learner, which must instantaneously decide whether perform check label or discard instance. approach inspired optimal experimental design theory iterative aspect decision-making process tackled setting threshold on informativeness points. evaluated using numerical simulations Tennessee Eastman Process simulator. results confirm that selecting examples suggested algorithm allows faster reduction prediction error.