The marine predator algorithm (MPA) often struggles with issues such as low population diversity, susceptibility to local optima, and population aggregation when optimizing XGBoost, which is sensitive to numerous parameters. To address these challenges, we propose the dual-population-iterating and prey-gathering marine predator algorithm (DPGMPA) and apply it to XGBoost parameter optimization. Initially, we utilize a dual-population-iterating strategy to create the original population and an opposition-based learning population. Brownian motion and Laplace distribution respectively control the step size to guide population updates toward the global optimum, enhancing population diversity. Subsequently, the preys-gathering strategy is employed, selecting prey with the highest accuracy in each iteration and having neighboring preys converge towards them to mitigate local optima. Furthermore, we optimize the Fish Aggregating Devices (FADs) by introducing variable weight control, addressing the population aggregation issue. To validate the parameter optimization capability of this algorithm for XGBoost, we conducted experiments on six datasets, using accuracy and F1 score as evaluation metrics, and compared the results with five other algorithms. The experimental results demonstrate that the DPGMPA algorithm more effectively escapes local optima when optimizing the XGBoost classifier, significantly improving XGBoost's classification performance.