M-GBDT2NN: A more generalized framework of GBDT2NN for online update

摘要

The large-scale, scalable, flexible characteristics make the researches of online update more important in the Internet of Things (IoT). Gradient Boosting Decision Tree (GBDT) is commonly used to deal with the numerical data, while it cannot be updated online. To better solve the prediction problems, we propose a more generalized approach, M-GBDT2NN, based on GBDT2NN. Compared with GBDT2NN, the new method is also applicable to the multi-classification problems, besides binary classification, regression. In the new framework, it takes the iteration of GBDT as the smallest unit to ensure the additive relation among the distilled models, and it predicts a probability vector rather than a numerical value. This paper analyzes the generalization and the ability of online update of M-GBDT2NN. The experimental results demonstrate that the proposed method can perform better than the other methods in both multi-classification problems and online update.