Stochastic Gradient Boosting

Gradient Boosting Machine (GBM) is a decision tree-based algorithm. Boosting combines multiple simple models (weak learners) into a single composite model. It is iteration-based and uses residual fitting to minimize a loss of function. Initially, a decision tree model is created (tree 1) to fit to the original data, and residuals are calculated from the fit. Then, another tree (tree 2) is built that fits to the residuals, and the residuals are calculated between the fit of tree 2 and the residuals from the initial fit, and another decision tree model (tree 3) is fit on that. In this manner, the decision trees are created sequentially, and the predictions of all the trees are added together to form the final prediction. Care is taken to not overfit to the data.

Use this algorithm when:

• Causal data is available.

• Multiple user defined causal variables exist and causal variables from the Trend Cloud are being used.