多示例多标记主动学习方法的研究及应用
发布时间：2019-05-09 13:21
【摘要】：主动学习作为一种能够解决传统分类问题中样本的标记信息缺失问题的机器学习方法受到了科研及实际应用领域的关注。多示例多标记学习作为一种新型的机器学习方法,能够处理现实任务中复杂的学习任务,目前尚未看到主动学习应用于多示例多标记学习任务中的研究。本文主要开展多示例多标记主动学习方法及其应用的研究,主要的工作内容如下:(1)设计了基于标记排序的多示例多标记主动学习方法框架MIMLAL。对传统主动学习模型、选择规则和多示例多标记框架等相关内容进行研究,根据目前较为先进的基于标记的机器学习思想提出了多示例多标记主动学习方法框架MIMLAL,解决主动学习任务计算量复杂信息容易丢失的问题。(2)提出两种具体的多示例多标记主动学习算法MIMLAL-A和MIMLAL-R。MIMLAL-A为基于标记空间最大均值差异的多示例多标记主动学习方法。传统的基于标记的分类方法使用关键示例的方法在主动学习中不够准确。本文依据最大均值差异MMD的思想将单示例主动学习价值函数LCI扩展到多示例下得到多示例MILCI并依此提出多示例多标记主动学习方法MIMLAL-A。MIMLAL-R为基于示例预测值排序的多示例多标记主动学习方法。MILCI使用平均的思想会引入不相关示例的无效信息,对此本文提出对示例按预测值排序加权的RLCI价值函数并针对单一选择规则对样本区分度不够的问题,基于结合RLCI和标记选择的新型选择规则提出MIMLAL-R方法。(3)实现了MIMLAL-A和MIMLAL-R方法在蛋白质功能预测问题上的应用。本文首先将新物种蛋白质功能预测任务抽象为多示例多标记主动学习问题,然后应用多示例多标记主动学习方法进行建模。在多个新物种的蛋白质功能预测任务中,通过对比发现,我们的多示例多标记主动学习方法取得了非常好的预测性能,其中MIMLAL-R在大多情况下性能最优。
[Abstract]:As a kind of machine learning method which can solve the problem of missing marker information of samples in traditional classification problems, active learning has attracted much attention in the field of scientific research and practical application. As a new machine learning method, multi-example multi-tag learning can deal with complex learning tasks in real-world tasks. At present, active learning has not been applied to the study of multi-example multi-tag learning tasks. The main contents of this paper are as follows: (1) A multi-sample multi-tag active learning method framework MIMLAL. based on tag sorting is designed in this paper, and the main work is as follows: (1) A multi-example multi-tag active learning method framework based on tag sorting is designed. In this paper, the traditional active learning model, selection rules and multi-example multi-tag framework are studied. According to the advanced markup-based machine learning idea, a multi-example multi-tag active learning method framework, MIMLAL, is proposed. To solve the problem that the complex information of active learning task computation is easy to lose. (2) two specific multi-example multi-tag active learning algorithms MIMLAL-A and MIMLAL-R.MIMLAL-A are proposed based on the maximum mean difference in tag space. Multi-example multi-tag active learning method. The traditional tag-based classification method using key examples is not accurate enough in active learning. In this paper, we extend the single example active learning value function LCI to the multi-sample MILCI according to the idea of the maximum mean difference MMD. Accordingly, we propose the multi-sample multi-marker active learning method MIMLAL-A.MIMLAL-R as the example-based active learning method. Multi-example and multi-tag active learning method for value ranking. MILCI uses the idea of average to introduce invalid information about unrelated examples. In this paper, the RLCI value function weighted according to the predicted value of the example is put forward, and the problem of insufficient classification of the sample area by a single selection rule is put forward. Based on the new selection rules combined with RLCI and marker selection, a MIMLAL-R method is proposed. (3) the application of MIMLAL-A and MIMLAL-R methods in protein function prediction is realized. In this paper, the task of protein function prediction of new species is abstracted as a multi-example multi-tag active learning problem, and then a multi-example multi-tag active learning method is used to model the problem. In the prediction task of protein function in many new species, we find that our multi-example multi-marker active learning method has achieved very good prediction performance, among which MIMLAL-R has the best performance in most cases.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181
本文编号：2472801
[Abstract]:As a kind of machine learning method which can solve the problem of missing marker information of samples in traditional classification problems, active learning has attracted much attention in the field of scientific research and practical application. As a new machine learning method, multi-example multi-tag learning can deal with complex learning tasks in real-world tasks. At present, active learning has not been applied to the study of multi-example multi-tag learning tasks. The main contents of this paper are as follows: (1) A multi-sample multi-tag active learning method framework MIMLAL. based on tag sorting is designed in this paper, and the main work is as follows: (1) A multi-example multi-tag active learning method framework based on tag sorting is designed. In this paper, the traditional active learning model, selection rules and multi-example multi-tag framework are studied. According to the advanced markup-based machine learning idea, a multi-example multi-tag active learning method framework, MIMLAL, is proposed. To solve the problem that the complex information of active learning task computation is easy to lose. (2) two specific multi-example multi-tag active learning algorithms MIMLAL-A and MIMLAL-R.MIMLAL-A are proposed based on the maximum mean difference in tag space. Multi-example multi-tag active learning method. The traditional tag-based classification method using key examples is not accurate enough in active learning. In this paper, we extend the single example active learning value function LCI to the multi-sample MILCI according to the idea of the maximum mean difference MMD. Accordingly, we propose the multi-sample multi-marker active learning method MIMLAL-A.MIMLAL-R as the example-based active learning method. Multi-example and multi-tag active learning method for value ranking. MILCI uses the idea of average to introduce invalid information about unrelated examples. In this paper, the RLCI value function weighted according to the predicted value of the example is put forward, and the problem of insufficient classification of the sample area by a single selection rule is put forward. Based on the new selection rules combined with RLCI and marker selection, a MIMLAL-R method is proposed. (3) the application of MIMLAL-A and MIMLAL-R methods in protein function prediction is realized. In this paper, the task of protein function prediction of new species is abstracted as a multi-example multi-tag active learning problem, and then a multi-example multi-tag active learning method is used to model the problem. In the prediction task of protein function in many new species, we find that our multi-example multi-marker active learning method has achieved very good prediction performance, among which MIMLAL-R has the best performance in most cases.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181
【参考文献】
相关期刊论文 前5条
1 吴建盛;;基于新型机器学习方法的蛋白质功能预测与分析[J];信息通信;2012年05期
2 孙吉贵;刘杰;赵连宇;;聚类算法研究[J];软件学报;2008年01期
3 龙军;殷建平;祝恩;赵文涛;;主动学习研究综述[J];计算机研究与发展;2008年S1期
4 赵悦;穆志纯;;基于委员会投票选择方法的主动学习的研究[J];太原理工大学学报;2006年04期
5 殷志祥;蛋白质结构预测方法的研究进展[J];计算机工程与应用;2004年20期
相关硕士学位论文 前1条
1 唐俊;基于多示例多标记学习的手机游戏道具推荐[D];南京大学;2015年
，本文编号：2472801
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2472801.html
