Miks ei saa lihtsalt õppida gradienttõstmise väljunditest?

In gradient boosting, training occurs on the residuals (gradients of the error), not on the outputs of previous models. If you try to train the next base algorithm solely on the outputs of the previous ones, the model will not correctly minimize the error because it does not consider the direction and magnitude of the error, only the final prediction.

Training on gradients allows sequential correction of previous models' errors, ensuring convergence and quality improvement. If training is only on outputs, the model may overfit or not reach the optimum, as it does not consider how strongly and in which direction the predictions need to be adjusted.

Example: in classical gradient boosting, the next regressor is trained on the difference between the true values and current predictions (residuals), not on the predictions themselves.