The alternating decision tree (ADTree) learning algorithm is a new method for classification that generalizes decision trees, voted decision trees, and voted decision stumps. It is designed to be more interpretable than traditional boosted decision trees like C5.0. ADTrees are represented as weighted votes of simple prediction rules, allowing for a more compact and interpretable classification rule. The algorithm uses boosting to learn ADTrees, and experimental results show that it is competitive with C5.0 and generates smaller, more interpretable rules. ADTrees also provide a natural measure of classification confidence, known as the classification margin, which can be used to improve accuracy by abstaining from predicting difficult examples. ADTrees are similar to option trees, which have been shown to improve classification error compared to single decision trees. The algorithm uses AdaBoost to learn ADTrees, and the resulting classifiers are structured as a sum of base rules. Each base rule is a simple prediction rule that contributes to the final classification. The algorithm maintains a set of preconditions and rules, and iteratively adds new rules to the tree. The classification is determined by the sign of the sum of the predictions from the base rules. ADTrees are more interpretable than traditional decision trees because the contribution of each decision node can be understood in isolation. This allows for easier interpretation of the classifier. The algorithm also provides a measure of confidence, the classification margin, which is the absolute value of the sum of the predictions. This measure can be used to improve accuracy by abstaining from predicting examples that are difficult to classify. Experiments show that ADTrees perform well on various data sets, often outperforming C5.0 and other methods. The algorithm is particularly effective on smaller data sets, where it achieves lower test error and smaller classifier size. The classification margin is also shown to be a reliable measure of confidence, with empirical evidence supporting its use in improving classification accuracy. The algorithm is also robust to noise and can be used to abstain from predicting on difficult examples, improving overall accuracy. Future work includes further research on using calibration graphs to decide when to abstain from prediction and integrating a more robust boosting algorithm.The alternating decision tree (ADTree) learning algorithm is a new method for classification that generalizes decision trees, voted decision trees, and voted decision stumps. It is designed to be more interpretable than traditional boosted decision trees like C5.0. ADTrees are represented as weighted votes of simple prediction rules, allowing for a more compact and interpretable classification rule. The algorithm uses boosting to learn ADTrees, and experimental results show that it is competitive with C5.0 and generates smaller, more interpretable rules. ADTrees also provide a natural measure of classification confidence, known as the classification margin, which can be used to improve accuracy by abstaining from predicting difficult examples. ADTrees are similar to option trees, which have been shown to improve classification error compared to single decision trees. The algorithm uses AdaBoost to learn ADTrees, and the resulting classifiers are structured as a sum of base rules. Each base rule is a simple prediction rule that contributes to the final classification. The algorithm maintains a set of preconditions and rules, and iteratively adds new rules to the tree. The classification is determined by the sign of the sum of the predictions from the base rules. ADTrees are more interpretable than traditional decision trees because the contribution of each decision node can be understood in isolation. This allows for easier interpretation of the classifier. The algorithm also provides a measure of confidence, the classification margin, which is the absolute value of the sum of the predictions. This measure can be used to improve accuracy by abstaining from predicting examples that are difficult to classify. Experiments show that ADTrees perform well on various data sets, often outperforming C5.0 and other methods. The algorithm is particularly effective on smaller data sets, where it achieves lower test error and smaller classifier size. The classification margin is also shown to be a reliable measure of confidence, with empirical evidence supporting its use in improving classification accuracy. The algorithm is also robust to noise and can be used to abstain from predicting on difficult examples, improving overall accuracy. Future work includes further research on using calibration graphs to decide when to abstain from prediction and integrating a more robust boosting algorithm.