Adaptive Batch Size Selection in Active Learning for Regression

Training supervised machine learning models requires labeled examples. A judicious choice of examples is helpful when there a significant cost associated with assigning labels. This article improves upon promising extant method – Batch-mode Expected Model Change Maximization (B-EMCM) for selecting to be regression problems. Specifically, it develops and evaluates alternate strategies adaptively batch size in B-EMCM.<br/> By determining the cumulative error that occurs from estimation stochastic gradient descent, stop criteria each iteration can specified ensure selected candidates are most beneficial model learning. new methodology compared B-EMCM via mean absolute root square over ten iterations benchmarked against data sets.<br/> Using multiple sets metrics across all methods, one variation AB-EMCM, max bound accumulated (AB-EMCM Max), showed best results an adaptive approach. It achieved better squared (RMSE) (MAE) than other nonadaptive methods while reaching result nearly same number as non-adaptive methods.