A company collected 10,000 equipment event samples over 3 months to build a predictive maintenance model. There are 100 failure events evenly spread across 50 machine types. How should the company prepare the training data to improve model accuracy given the class imbalance and sparse failures per type?

SMOTE is the best choice because it synthesizes new, realistic failure samples based on existing ones, effectively increasing the representation of the minority class (failures) without simply duplicating data. This addresses both the class imbalance and the sparsity of failures per machine type, providing the model with more examples to learn from. Adjusting class weights helps with imbalance but doesn't create new data to address sparsity. Undersampling non-failure events would reduce the overall dataset size, potentially losing valuable information, and applying SMOTE to non-failure events is incorrect as SMOTE is designed for minority classes.