代价敏感的主动学习研究

Traditional supervised learning usually requires a large amount of labeled data to train the model, leading to high cost of human annotation. Active learning selectively queries the most valuable information from oracles, and can significantly reduce the cost. However, existing studies on active learning evaluate the cost simply with the number of queried examples, and thus cannot handle the diverse costs from aspects of examples. This project proposes to study on this important issue, and mainly includes the following research contents: for sensitiveness to the querying cost, proposing an active learning approach to deal with diverse annotation difficulties among examples. This project studies the cost-sensitiveness of active learning, and tries to apply the methods to task of risk prediction for news events. It is expected to publish 2-3 high quality papers on important journals and conferences, apply 1 patents.

传统监督学习往往基于大量已标记数据进行模型训练，导致高昂的人工标注代价。主动学习选择性地向专家查询部分最有价值信息，可有效减少学习代价。然而，现有方法简单地将代价等同为标注样本数目，无法有效处理实际应用中在样本等方面存在的代价差异。本项目针对这一问题进行研究，主要内容包括：针对样本查询代价敏感性，提出适应样本多样化标注难度的主动学习方法。本项目针对性的研究主动学习中的代价敏感性，并围绕新闻事件风险预测任务进行应用研究。成果有望在国内外重要学术期刊和会议上发表高质量论文2-3篇，申请专利1项。

传统监督学习往往基于大量已标记数据进行模型训练，导致高昂的人工标注代价。主动学习选择性地向专家查询部分最有价值信息，可有效减少学习代价。然而，现有方法简单地将代价等同为标注样本数目，无法有效处理实际应用中在样本等方面存在的代价差异。为了解决这一问题，本项目针对主动学习中代价敏感性，提出了一种能够适应样本多样化标注难度的主动学习方法，并成功将这一方法应用于新闻事件风险预测任务。