主动学习初始化策略及停止准则研究

Currently, many machine learning tasks face a practical problem that we can easily acquire a large number of unlabeled instances, but labeling these instances is quite burdensome and expensive. Active learning aims to query as little data as possible while maximize the learning performance by using a selective sampling strategy. The initialization strategy and stopping criterion for active learning are two factors that heavily influence the effect of active learning. Since current studies about these two aspects still have a lot of weak points, based on our previous works on active learning, this project explores the corresponding theoretical model and technical methods. First, we attempt to explore a new active initialization strategy that uses a nearest neighbor criterion to select the most representative instances and construct the initial labeled dataset. Secondly, we try to investigate the influence of active initialization strategy to the active learning algorithms employed in subsequent steps. Comparative experiments will be constructed to verify whether initialization strategy is able to directly affect the effectiveness of active learning methods. Lastly, we will search for a new stopping criterion for active learning which does not require a separate validation set. The average retraining-based variance will be used as the stopping criterion. This project will be of significance for promoting the theoretical analysis and practical application of active learning and bringing down the human annotation cost.

当前,很多机器学习任务面临的一个现实问题是：未标记样本数量众多，人工标注成本高昂。主动学习旨在通过交互迭代式的选择性采样，以尽可能少的标注成本来获得性能较高的分类模型。主动学习的初始化策略和停止准则是影响其实际应用效果的两个重要因素。针这两方面的研究尚存在诸多不足，本项目在前期研究基础上，探索了相应的理论问题和技术方法。首先，研究了一种新的主动学习初始化策略，该策略通过一种最近邻准则来寻找最具代表性的样本构建初始训练集；然后，探索了初始化策略对主动学习算法有效性的影响，设计对比实验来验证初始化策略能否直接影响主动学习算法的效益；最后，探索一种新的不依赖独立验证集的主动学习停止准则，该准则采用平均重训练方差作为停止指标。本项目的预期成果对推进主动学习的基础理论研究及其实用化，减少人工标注成本具有重要的现实意义。

本项目主要研究了主动学习初始化策略和停止准则，对上述两个问题的攻坚研究能够促进主动学习在实际生活中的应用，有效地节省人工标注成本，具有重要的现实意义。本项目围绕研究任务展开，主要内容包括：1）研究了主动学习初始化算法，提出了基于最近邻距离的主动学习初始化方法，该算法能够找到一个最具代表性的初始数据集，即该数据集与剩余的未标记数据之间的整体相似性很高。2）研究了初始化策略对主动学习算法的影响，发现初始化算法对整个主动学习样本标注过程影响很大，好的初始化算法更有可能提升主动学习算法的整体性能，性能不佳的初始化算法很可能会降低性能。3）研究了主动学习停止准则及评价指标，提出了基于重训练方差的主动学习停止准则及评价指标，帮助找到人工标记成本与机器学习模型准确率的平衡点。此外，基于此项目，协助培养2名博士研究生，发表3篇SCI论文，预备出版专著一本。本项目研究填补了相关领域的空白，能够推动主动学习在不同领域的应用。

内容获取失败，请点击重试