Multi-Classification Data Stream Algorithm Based on One-Vs-Rest Strategy.

Abstract: In the data stream learning scenario, the whole picture of the data can't be observed, and the data may change dynamically, thus increasing the complexity and imbalance of the data. Aiming at the characteristics of data class changes (appearance, disappearance, and reappearance) of multi-class data stream, a Matthews Adaptive XGBoost algorithm based on One-Vs-Rest strategy is proposed. For the three cases of class change, the One-Vs-Rest strategy is used to build the model, and a binary classification model is created for each class. Aiming at the problem that the model gradually forgets the knowledge of the class after the class disappears, a management mechanism based on class frequency is proposed. In view of the problem that the restarted model affects the overall performance of the ensemble after the class is reproduced, the sliding window is used to initialize it. Aiming at the problem of ensemble classifiers, a mechanism to adapt the number of base classifiers is proposed. Experiments on real and synthetic datasets show that the improved algorithm improves Kappa and PMAUC indicators by 1.03% and 0.82%, respectively. CCS CONCEPTS • Computing methodologies • Machine learning • Machine learning algorithms