1. Initial Prediction: You start with an initial, very simple model. For regression tasks, this is often just the average (mean) of the target variable for all training instances. For binary classification, it might be the log-odds of the positive class. This provides a foundational, albeit crude, prediction.
2. Calculate Residuals (Errors): For each training instance, you calculate the "residual." This is simply the numerical difference between the actual target value and the current ensemble's cumulative prediction. These residuals are effectively the errors that the models built so far are making, and they represent the "unexplained" part of the target variable.
3. Train New Learner on Residuals: A new base learner (almost always a shallow decision tree, often specifically called a "regression tree" because its goal is to predict these numerical errors or residuals) is trained. Its target variable is not the original target value, but rather these residuals calculated in the previous step. The new tree's objective is to learn patterns in these errors and predict them.
4. Add to Ensemble (with Learning Rate): The prediction of this newly trained base learner (which is essentially its predicted residual) is then added to the ensemble's current cumulative prediction. However, it's added with a small multiplier called the learning rate (also known as "shrinkage"). This learning rate is a crucial hyperparameter that controls the step size or the contribution of each new tree to the overall ensemble. Adding a small learning rate helps prevent the model from quickly overfitting and allows for a smoother, more controlled learning process.
5. Iterative Process: Steps 2-4 are repeated for a specified number of iterations (which is equivalent to the number of trees in the ensemble). In each iteration, a new tree is trained to predict the remaining errors of the combined predictions from all the previous trees. This iterative process gradually reduces the overall error of the entire ensemble.
6. Final Prediction: The final prediction for a new, unseen instance is the sum of the initial prediction and the scaled predictions of all the individual base learners (trees) that were added to the ensemble during the training process.

● Highly Accurate and Flexible: GBM is incredibly powerful and consistently achieves state-of-the-art results on structured (tabular) data across a wide range of problems, for both classification and regression.

● Versatile: It can handle various types of data and is very flexible in adapting to different prediction tasks.

● Robustness with Proper Tuning: When carefully tuned with appropriate hyperparameters, GBM models are very robust and can generalize exceptionally well to unseen data.

● Prone to Overfitting: Because it aggressively tries to fit the training data by reducing residuals, GBM can be prone to overfitting if its hyperparameters are not tuned properly (e.g., if there are too many trees, if the learning rate is too high, or if individual trees are too deep).

● Computationally Intensive and Sequential: The sequential nature of its training means that it can be slower to train compared to bagging methods, especially with a very large number of trees or complex datasets.

● More Complex to Tune: It generally has a larger number of hyperparameters that need careful tuning for optimal performance, which can require more expertise and computational resources.