Short answers, pulled from the story.
An alternating decision tree is a machine learning method for classification that generalizes decision trees and connects to boosting. It consists of alternating decision nodes, which specify a predicate condition, and prediction nodes, which contain a single number. An instance is classified by following all paths where decision nodes are true and summing the prediction nodes traversed.
ADTrees were introduced by Yoav Freund and Llew Mason. The original publication contained typographical errors, and later clarifications and optimizations were provided by Bernhard Pfahringer, Geoffrey Holmes, and Richard Kirkby.
In binary classification trees like CART and C4.5, an instance follows only one path through the tree. In an ADTree, an instance follows all paths for which all decision nodes are true, and the final classification is the sum of prediction node values along those paths.
Implementations of the alternating decision tree algorithm are available in Weka and JBoost.
Instances that are misclassified during training are given a larger weight in the next iteration, while accurately classified instances receive reduced weight. This re-weighting focuses successive rules on the examples the current model gets wrong.
The sign of the final summed score determines the classification, while the magnitude indicates confidence in the prediction. The original authors also identify three levels of interpretation: individual nodes, sets of nodes on a shared path, and the tree as a whole.