This paper presents an empirical comparison of voting classification algorithms, including Bagging, AdaBoost, and Arc-x4, using decision tree inducers and Naive-Bayes. The study aims to understand how these algorithms affect classification error through a bias-variance decomposition. Key findings include: 1. **Bagging**: Reduces variance for unstable methods, improving performance. 2. **AdaBoost and Arc-x4**: Reduce both bias and variance for unstable methods but increase variance for Naive-Bayes, which is very stable. 3. **Pruning and Probabilistic Estimates**: Disabling pruning and using probabilistic estimates (scoring) can further reduce error for decision trees. 4. **Wagging and Backfitting**: Wagging adds Gaussian noise to weights, while backfitting refines probability estimates by reusing the original training set. 5. **Mean-Squared Error (MSE)**: Voting methods significantly reduce MSE compared to non-voting methods. The study uses a large dataset set with over 1000 instances and four datasets with over 10,000 instances to ensure reliability. The results highlight the effectiveness of voting algorithms in improving classification accuracy and provide insights into their underlying mechanisms.This paper presents an empirical comparison of voting classification algorithms, including Bagging, AdaBoost, and Arc-x4, using decision tree inducers and Naive-Bayes. The study aims to understand how these algorithms affect classification error through a bias-variance decomposition. Key findings include: 1. **Bagging**: Reduces variance for unstable methods, improving performance. 2. **AdaBoost and Arc-x4**: Reduce both bias and variance for unstable methods but increase variance for Naive-Bayes, which is very stable. 3. **Pruning and Probabilistic Estimates**: Disabling pruning and using probabilistic estimates (scoring) can further reduce error for decision trees. 4. **Wagging and Backfitting**: Wagging adds Gaussian noise to weights, while backfitting refines probability estimates by reusing the original training set. 5. **Mean-Squared Error (MSE)**: Voting methods significantly reduce MSE compared to non-voting methods. The study uses a large dataset set with over 1000 instances and four datasets with over 10,000 instances to ensure reliability. The results highlight the effectiveness of voting algorithms in improving classification accuracy and provide insights into their underlying mechanisms.