基于多样化查询的多标记主动学习研究

Multi-label learning is an important research area of machine learning, where each object is simultaneously associated with multiple labels. Manually annotating a large set of multi-label objects will lead to high cost. Active learning, which selectively queries the most important supervision information from the oracle, can reduce the labeling cost significantly. Most existing methods query all labels of one instance at a time. Such a simple strategy is less effective, and cannot meet the varying requirements of the learning model on the supervision information. This project proposes to study on this important issue, and the research contents include: 1) proposing a multi-label active learning (MLAL) approach by querying the relevance on instance-label pairs; 2) proposing a MLAL approach by querying the relevance ordering of label pairs; 3) analyzing the varying requirements of the learning model and building the theoretical basic for it; 4) designing a MLAL approach to adaptively incorporate different query types. Our project provides a novel direction to further improve the performance of MLAL. It is expected to publish 4-6 papers on high quality journals or conferences and apply 1-2 patents.

多标记学习是机器学习中的一个重要研究方向，其学习对象同时与多个标记相关联，人工标注大量样本将导致高昂代价。主动学习通过选择性地向用户查询部分最有价值的监督信息，可有效减少用户标注代价。现有方法往往采用单一的查询方式向用户询问一个样本的所有标记，效率较低，且无法满足模型在学习过程中随监督信息增加而不断变化的需求。本项目针对这一问题进行研究，主要内容包括：（1）提出基于“示例-标记”对相关性查询的多标记主动学习方法；（2）提出基于标记间相关度排序查询的多标记主动学习方法；（3）分析多标记主动学习中各阶段模型对监督信息的需求变化，并构建其理论基础；（4）设计自适应融合多种查询方式的多标记主动学习方法。本项目为进一步提高多标记主动学习效果开辟了新的思路和方向，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4-6篇，并申请专利1-2项。

多标记学习是机器学习中的一个重要研究方向，其学习对象同时与多个标记相关联，人工标注大量样本将导致高昂代价。主动学习通过选择性地向用户查询部分最有价值的监督信息，可有效减少用户标注代价。现有方法往往采用单一的查询方式向用户询问一个样本的所有标记，效率较低，且无法满足模型在学习过程中随监督信息增加而不断变化的需求。本项目针对这一问题进行研究，提出基于“示例-标记”对相关性查询的多标记主动学习方法；提出基于标记间相关度排序查询的多标记主动学习方法；提出自适应融合多种查询方式的多标记主动学习方法。发表高水平论文10篇，申请专利3项，培养研究生4名。