Abstract:

In the real world, the available training data is usually small and easily outdated, so a large number of new data sets need to be continuously collected and labelled. To address this problem, a transfer learning classification method based on SAMME and TrAdaBoost algorithms is proposed. The core idea of the method is to filter useful samples from the old data set to help the model identify new video samples, where the feature distributions of the old and new video traffic datasets are different. At the same time, this method combines the SAMME algorithm to extend the TrAdaBoost algorithm from only two classifications to multi-classification. Experimental results show that the proposed method can better achieve fine-grained classification of six types of video traffic and reduce the waste of a large number of old data sets compared to existing methods.

Key words: transfer learning, video services, network traffic classification

摘要：

在现实世界中，可用的训练数据通常较少，且很容易过时，所以需要不断采集和标记大量新的数据集；针对此问题，提出一种基于SAMME和TrAdaBoost算法的迁移学习分类方法。该方法的核心思想是：从老视频流数据集中筛选出有用的样本来帮助模型识别新的未知视频流集样本，这里新老视频流数据集的样本特征分布是不相同的。同时该方法结合SAMME算法将TrAdaBoost算法从只可实现两分类扩展至多分类。实验结果表明，与现有方法比较，该方法能更好地实现对六种类型视频流的精细分类，并减少大量已标注老数据集的浪费。

关键词: 迁移学习, 视频业务, 网络流分类