Boosting is an ensemble learning meta-algorithm primarily used in machine learning to reduce bias and variance. It sequentially builds a strong predictive model from a series of weak learners, with each new learner focusing on the errors of its predecessors.