A team trains 10 decision tree classifiers, each on a different random subset of training data (with replacement). They combine predictions by majority vote. What is this ensemble technique, and what property of the base models is essential for this approach to improve over a single tree?