Because of the complex mechanisms in pulsed disk and doughnut columns (PDDCs), traditional empirical functions often fail to make accurate predictions in new datasets, such as different experimental conditions or different PDDC structures, indicating a lack of generalizability. In this work, some machine learning techniques such as random forest regression (RFR), least absolute shrinkage and selection operator, support vector regression (SVR), and artificial neural network are developed to predict dispersed phase holdup based on experimental data collected from numerous studies. Two training methods were used: One is to randomly divide the collected data into groups for training and testing, and the other is to separate the data of one study for testing and training in data from other studies. These methods were used to compare and analyze the accuracy, generalizability, and stability of these models, using the mean relative error (MRE) as the performance evaluation criterion. SVR has an MRE of 15.0% in the test set and 11.0% in the entire dataset, outperforming other alternative models in both efficiency and ability to mitigate overfitting. Furthermore, the relative importance of each parameter in influencing holdup was analyzed by RFR.