Investigating Query by Committee and Inverse‐Distance Based Exploration for Active Learning as Learning Strategies: An Application in Wood Dating

ABSTRACT Fourier transform infrared (FTIR) spectroscopy, a non‐destructive method, is capable of measuring changes in the major wood components cellulose, hemicellulose and lignin during degradation. Previous research evaluated the application of supervised machine learning based on molecular decay over specific preservation conditions, achieving a root mean square error (RMSE) of around 100 years and covering an 800‐year period. These molecular decay dating models depend on the availability of labels, which were dated dendrochronologically. Active learning reduces the labeling effort by querying the most informative samples. The active learning approaches Query by Committee (QBC) and inverse‐distance based weighting (IDW) were assessed for support vector regression (SVR) and random forest (RF) as the underlying prediction model. Results from several replicated runs with 200 queries each show that both methods effectively reduce RMSE compared to random sampling but vary depending on the underlying prediction model type. QBC and IDW showed similar improvements of performance with RF by reducing the RMSE by 6 years. IDW significantly enhances SVR performance, reducing the RMSE by 10 years compared to the 1‐year improvement with QBC. Active learning related parameters such as initial sample size and committee size besides initial data set variation impact the outcomes of these active learning approaches. Both strategies outperform supervised machine learning baseline models, demonstrating the effectiveness of active learning in enhancing model accuracy.