Using data-driven algorithms to accurately forecast solar flares requires reliable data sets.The solar flare dataset is composed of many non-flaring samples with a small percentage of flaring samples.This is called the class imbalance problem in data mining tasks.The prediction model is sensitive to most classes of the original data set during training.Therefore,the class imbalance problem for building up the flare prediction model from observational data should be systematically discussed.Aiming at the problem of class imbalance,three strategies are proposed corresponding to the data set,loss function,and training process:Type Ⅰ resamples the training samples,including oversampling for the minority class,undersampling,or mixed sampling for the majority class.Type Ⅱ usually changes the decision-making boundary,assigning the majority and minority categories of prediction loss to different weights.Type Ⅲassigns different weights to the training samples,the majority categories are assigned smaller weights,and the minority categories are assigned larger weights to improve the training process of the prediction model.The main work of this paper compares these imbalance processing methods when building a flare prediction model and tries to find the optimal strategy.Our results show that among these strategies,the performance of oversampling and sample weighting is better than other strategies in most parameters,and the generality of resampling and changing the decision boundary is better.