Umeå universitets logga

Using signal detection theory, we meta-analyzed existing research testing the classification performance of supervised machine learning (ML) models applied in human resource (HR) contexts. Our meta-analysis contained 6605 effect sizes cross-nested by study (N1 = 249) and unique dataset (N2 = 152). We conducted separate cross-classified multilevel modeling analyses predicting six different classification performance indices. We tested hypotheses regarding the effects of algorithm type, sample size, number of predictors used, the number of outcome classes, and outcome class imbalance on model classification performance. Boosting and random forest algorithms performed best across classification performance indices. However, both come at relatively great computational expense, and solutions can be difficult to explain. Decision trees were the best-performing algorithms with relatively lower computational expense and easily interpretable solutions. We found limited evidence for the effect of sample size on classification performance. We found stronger support suggesting that using more predictors to train machine learning models results in better classification performance, but only when those predictors have substantive value. Finally, we found strong evidence that outcome class imbalance influences classification performance indices differently. More imbalanced outcome classes are associated with higher accuracy scores, which can be misleading, and lower precision and F1 scores, which may be better estimates of true classification performance. Our findings provide valuable insights into developing ML tools better suited to support HR functions.