In domains dealing with unbalanced data categorization, the dataset is expected to be balanced by adding a few more classes of data, which results in more data and low data quality. Therefore, The paper proposes an innovative hybrid method that combines game-based mixed sampling and an improved genetic algorithm to solve imbalanced data classification with a small dataset. Initially, the framework utilizes game theory to establish balanced sampling methods and ratios to address data imbalances. Subsequently, it employs the SelectKBest technique to optimize feature selection. Finally, the improved genetic algorithm will refine the sampling size and sample selection. In the feature encoding stage of the genetic algorithm, an ensemble learning method is adopted, using K-nearest neighbor (KNN), decision tree (DT), support vector machine (SVM), random forest (RF), and Adaboost, combined with precision, F1 score, and MCC performance evaluation indicators to measure, preventing premature convergence and optimizing the entire solution space, thus enhancing data sampling quality. Determination of the minimum stable population size employs a sliding standard deviation approach. Empirical findings corroborate the efficacy of this approach in tackling challenges associated with imbalanced data classification, refining the sample space, and improving sample quality. This methodology demonstrates significant practical utility in augmenting classifier performance when dealing with imbalanced datasets.