大规模云平台任务终止状态的预测是云资源调度策略优化的关键步骤。本文以Google云平台的计算调度系统Borg为对 象进行研究，针对任务的各种终止状态极度不均衡和类重叠的问题，提出了一种类重叠度区分的自定义步长‐梯度提升决策树 （SP‐GBDT）任务终止状态预测方法，对任务终止状态进行细粒度的多分类预测，提高少数类任务状态的预测准确率。首先将终 止状态的多个类别拆分成若干个二类组合，通过支持向量数据描述模型（SVDD）筛选出类重叠度较低的最优二类组合。然后， 分别对最优的二类组合进行扩展采样比例的自定义步长欠采样。最后构建梯度提升决策树模型，将欠采样之后的数据进行多 分类。在Google云平台的运行监控日志数据集上进行验证，通过对比预测结果和预测过程的可解释性分析，SP‐GBDT模型能 够很好地降低数据集的不均衡比例以及类重叠的程度。与决策树和随机森林等常用多分类预测方法相比，所提算法的F1‐score 分别提高了30.39%和18.26%。

Task termination statuses prediction is one of the key technologies to realize resource scheduling optimization in large‑scale cloud platform. In this paper, using the Google Cloud Platform′s computing scheduling system Borg as the object, aiming at the problem that various termination statuses of tasks are extremely unbalanced and overlapping, a Self‑Paced ‑Gradient Boost Decision Tree (SP‑GBDT) is proposed to predict task termination statuses. Fine‑grained multi‑classification prediction of task termination state is carried out to improve the accuracy of predicting a few classes of task states. First, the termination state is divided into several two‑class combinations, and the best two‑class combination with low class overlap is selected by Support Vector Data Description model (SVDD). Then, the optimal combination of the two classes is under‑sampled for the custom step of the extended sampling scale. Finally, a gradient lifting decision tree model is constructed to classify the under‑sampled data. Verification on the operation monitoring log data set of the Google cloud platform, the experimental results show that the SP‑GBDT model can reduce the imbalance ratio of the dataset and the degree of class overlap through the study of the interpretability of the prediction process and results. Compared with the commonly used multi‑classification prediction methods such as decision trees and random forests, the F1‑score of the proposed algorithm increased by 30.39% and 18.26%, respectively.