基于实例的归纳式迁移学习研究

发布时间：2018-12-28 18:21

【摘要】：在全球信息化大潮的推动下,数据量日益增长,其增速让人惊叹,面对如此之多的信息和数据突如其来的冲击大脑,人们没有时间看这些数据,人们的关注已经转到可贵的应付手段上。因此,人们必须找到有效方法,应对处理大量的信息以此来高效、准确的找到有价值的信息。首先,本文针对具有相关性的数据,考虑数据之间具有层次关系,将与目标领域相关的数据根据相关性按不同比例调整权重,在经典迁移学习算法Tr Ada Boost中引入层次相关性、权重调整策略、错分样本权重约束,从而提出一种结合层次相关性的迁移学习算法,以解决具有层次关系的数据,高效调整权重的问题。在实验部分,本文将基于层次相关性的迁移学习算法与SVM、Tr Adaboost算法的正确率、查准率、查全率进行对比,对实验数据分析得到基于层次相关性的数据具有更准确的分类结果。针对目标领域与多个源领域都相近,但源领域与目标领域不一样,多个源领域也都不一样,都是一个大的源领域下面的多个小的源领域领域,这些小的源领域还有些相关性,从多个小的源领域往一个目标领域迁移。针对具有多个源领域,单个源领域样本不足,负迁移的问题,提出一种多源实例迁移学习算法,考虑多个源领域的知识,使目标领域能综合考虑运用每个源领域的知识,该方法先对源领域和目标领域做并集训练出分类器,经过测试后保留那些取并集后提升分类效果的源领域。然后,再将留下的源领域做并集,做Tr Adaboost迁移学习,测试后按照指定的规则选择最终的集合作为源领域,与目标领域共同训练出分类器。最后,通过实验对第三章,第四章所提出的基于层次相关性的迁移学习和多源实例迁移学习进行详细的描述,并将实验结果做出客观的分析总结,对比改进前后的正确率,查准率,查全率。证明了基于层次相关性的迁移学习算法和多源实例迁移学习算法能够非常有效地提高分类器的分类效果。
[Abstract]:Driven by the tide of global information technology, the amount of data is growing day by day, and the growth rate is amazing. In the face of so much information and data suddenly coming to the brain, people do not have time to look at these data. Attention has shifted to valuable coping techniques. Therefore, people must find an effective way to deal with a large amount of information in order to find valuable information efficiently and accurately. First of all, considering the hierarchical relationship between the data, this paper adjusts the weight of the data related to the target domain according to the correlation according to different proportions, and introduces the hierarchical correlation into the classical migration learning algorithm Tr Ada Boost. In order to solve the problem of data with hierarchical relationship and adjust weights efficiently, a transfer learning algorithm combining hierarchical correlation is proposed. In the experiment part, we compare the transfer learning algorithm based on hierarchical correlation with the correct rate, precision rate and recall rate of SVM,Tr Adaboost algorithm, and get more accurate classification results by analyzing the experimental data. The target domain is similar to multiple source domains, but the source domain is not the same as the target domain, and many source domains are not the same, they are many small source fields under a large source domain, and these small source domains have some relevance. Migrate from multiple small source domains to one target domain. In order to solve the problem of multiple source domains with insufficient samples and negative migration, a multi-source instance transfer learning algorithm is proposed, in which the knowledge of multiple source domains is considered, so that the target domain can comprehensively consider the use of knowledge in each source domain. The method firstly combines the source domain and the target domain to train the classifier. After the test, the source domain that is used to improve the classification effect is reserved. Then, the source domain is merged and the Tr Adaboost migration learning is done. After testing, the final set is selected as the source domain according to the specified rules, and the classifier is trained together with the target domain. Finally, the transfer learning based on hierarchical correlation and multi-source instance transfer learning proposed in chapter 3 and chapter 4 are described in detail through experiments, and the experimental results are objectively analyzed and summarized, and the correct rate before and after the improvement is compared. Precision rate It is proved that the hierarchical correlation based migration learning algorithm and the multi-source instance migration learning algorithm can effectively improve the classification effect of the classifier.
【学位授予单位】：辽宁大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【参考文献】