How algorithmic innovations drive performance

Random forests can be trained in parallel by growing individual trees on independent bootstrap samples. In contrast, the sequential approach of gradient boosting slows down training, which in turn complicates experimentation with a large number of hyperparameters that need to be adapted to the nature of the task and the dataset.

To expand the ensemble by a tree, the training algorithm incrementally minimizes the prediction error with respect to the negative gradient of the ensemble's loss function, similar to a conventional gradient descent optimizer. Hence, the computational cost during training is proportional to the time it takes to evaluate the impact of potential split points for each feature on the decision tree's fit to the current gradient.

..................Content has been hidden....................

You can't read the all page of ebook, please click here login for view all page.
Reset
44.192.38.143