In the context of using Amazon SageMaker to build a sentiment analysis model, your machine learning team needs to handle an unbalanced dataset where negative sentiment samples significantly outnumber positive ones. Which technique should be applied to address this issue effectively and improve model performance?