Breast cancer prediction datasets are usually class imbalanced, where the number of data samples in the malignant and benign patient classes are significantly different. Over-sampling techniques can be used to re-balance the datasets to construct more effective prediction models. Moreover, some related studies have considered feature selection to remove irrelevant features from the datasets for further performance improvement. However, since the order of combining feature selection and over-sampling can result in different training sets to construct the prediction model, it is unknown which order performs better. In this paper, the information gain (IG) and genetic algorithm (GA) feature selection methods and the synthetic minority over-sampling technique (SMOTE) are used for different combinations. The experimental results based on two breast cancer datasets show that the combination of feature selection and over-sampling outperform the single usage of either feature selection and over-sampling for the highly class imbalanced datasets. In particular, performing IG first and SMOTE second is the better choice. For other datasets with a small class imbalance ratio and a smaller number of features, performing SMOTE is enough to construct an effective prediction model.