关键词: 离群检测, 主动学习, 模型集成

Abstract:

Outlier detection tasks usually lack available labeled data, and outlier data only accounts for a small part of the entire data set. Compared to other data mining tasks, outlier detection is more difficult, and there is no single algorithm suitable for all the scenes. Therefore, combined with the idea of diversity model ensemble and active learning, this paper proposes an outlier ensemble detection method named Outlier Mining based on Active Learning（OMAL）. Under the guidance of the active learning framework, five unsupervised models based on statistics, similarity and axis-parallel subspace are selected as the base learners according to the comparative analysis of various learners. Then, the outlier and normal boundary data of each base learner are integrated, filtered and presented to the human experts for labeling to maximize information feedback from the human experts. Sampling from the labeled dataset and the dataset generated by the voting of the base learners. A supervised binary classification model based on Gradient Boosting Machine（GBM） is trained and applied to the full dataset to mining the final results. Experiments show that the AUC of OMAL method has been significantly improved while providing good performance and practical value.

Key words: outlier detection, active learning, model ensemble