A clustering-based active learning method to query informative and representative samples
A clustering-based active learning method to query informative and representative samples
复制标题
DOI：
10.1007/s10489-021-03139-y
复制
发表时间：
2022-02
影响因子：
5.3
通讯作者：
Xuyang Yan;Shabnam Nazmi;Biniam Gebru;Mohd M. Anwar;A. Homaifar;M. Sarkar;Kishor Datta Gupta
中科院分区：
文献类型：
--
作者：
Xuyang Yan;Shabnam Nazmi;Biniam Gebru;Mohd M. Anwar;A. Homaifar;M. Sarkar;Kishor Datta Gupta
主动学习（AL）已被广泛用于解决标记数据集的短缺。然而，大多数人工智能技术需要一个初始的一组标记的数据作为知识库来执行主动查询。初始标记集的信息量会显著影响后续的主动查询，从而影响主动学习的性能。在本文中，一个新的基于聚类的主动学习框架，即主动学习使用基于聚类的采样（ALCS），提出了同时考虑样本的代表性和信息性，使用没有先验标签信息。一个基于密度的聚类方法来探索集群结构的数据，而不需要详尽的参数调整。一个简单而有效的基于距离的查询策略被用来调整基于中心和基于边界的主动学习选择之间的采样权重。提出了一种新的基于双聚类边界的样本查询方法，在相邻聚类的边界上选择最不确定的样本。此外，我们开发了一个有效的多样性探索策略，以解决查询样本之间的冗余。我们广泛的实验提供了ALCS方法与最先进的方法的比较，表明ALCS比最先进的方法产生统计学上更好或相当的性能。
Active learning (AL) has widely been used to address the shortage of labeled datasets. Yet, most AL techniques require an initial set of labeled data as the knowledge base to perform active querying. The informativeness of the initial labeled set significantly affects the subsequent active query; hence the performance of active learning. In this paper, a new clustering-based active learning framework, namely Active Learning using a Clustering-based Sampling (ALCS), is proposed to simultaneously consider the representativeness and informativeness of samples using no prior label information. A density-based clustering approach is employed to explore the cluster structure from the data without requiring exhaustive parameter tuning. A simple yet effective distance-based querying strategy is adopted to adjust the sampling weight between the center-based and boundary-based selections for active learning. A novel bi-cluster boundary-based sample query procedure is introduced to select the most uncertain samples across the boundary among adjacent clusters. Additionally, we developed an effective diversity exploration strategy to address the redundancy among queried samples. Our extensive experimentation provided a comparison of the ALCS approach with state-of-the-art methods, exhibiting that ALCS produces statistically better or comparable performance than state-of-the-art methods.