Mitigations for Non-IID labels in Federated XGBoost Classification

Federated learning, a distributed collaborative framework, has become popular because it ensures participants' datasets stay local, which fundamentally alleviates privacy issues. However, unrestricted participants inevitably bring problems of non-independent and identically distributed data. These problems will cause non-convergence and poor generalization in the classification. This thesis provides a selection of the alternative classification matrix and a new objective function to adjust binary class imbalanced classification. It proposed an index for multi-class imbalance and an algorithm for hyperparameters selection. In addition, two sliding algorithms are proposed to improve the accuracy and convergence of the model by combining two objective functions' advantages.