Sensor-based human activity recognition (HAR) has important research value and significance in the field of health care. Previous researches on classification and recognition of sensor human activities have not considered the imbalance between different categories of behavior data. In order to solve the problem that the imbalance between data of different behavioral categories affects the accuracy of the algorithm, our algorithm uses the downsampling method to randomly extract two sets of data from large and small data sets, which are equal in number, and transform multiple imbalanced data into balanced data. Secondly, multiple weak classifiers are trained on multiple balanced datasets. Then, the algorithm takes the negative correlation and prediction accuracy of the weak classifier as the cost function and uses genetic algorithm to select the weak classifier which can make the highest value of the cost function to form the integrated classifier. The weak learner in the ensemble algorithm has high prediction accuracy and diversity. Finally, the algorithm uses the selected weak learners to construct an ensemble learner to classify human behavior. The experimental results show that the proposed algorithm can effectively improve the accuracy of unbalanced behavior recognition compared with the traditional algorithms.