稀疏分层概率自组织图实例迁移学习方法

Instance transfer learning model based on sparse hierarchical probabilistic self-organizing graphs

针对基于实例的迁移学习在关联多源异构领域数据时遇到的数据颗粒度不匹配问题,以单领域分层概率自组织图(Hi PSOG)聚类方法为基础,提出一种具有迁移学习能力的稀疏化非监督分层概率自组织图(TSHi PSOG)方法。首先,在源领域和目标领域分别基于概率混合多变量高斯分布生成分层自组织模型以便在多领域中分别提取不同粒度的表示向量,并用稀疏图方法通过概率准则控制模型增长;其次,利用最大信息系数(MIC),在具有富信息的源领域中寻找与目标领域表示向量最相似的表示向量,并利用这些源领域表示向量的类别标签细化目标领域数据分类;最后,在国际通用分类数据集20新闻组数据集和垃圾邮件检测数据集上进行了实验,结果表明算法可以利用源领域的有用信息辅助目标领域的分类问题,并使分类准确率最高提高约15.26%和9.05%;对比其他经典迁移学习方法,通过稀疏分层可以挖掘不同颗粒度的表示向量,分类准确率最高提高约4.48%和4.13%。

The current study of instance-transfer learning suffers from the mismatch between the granularities of data from multi-source heterogeneous domains. A Transfer Sparse unsupervised Hierarchical Probabilistic Self-Organizing Graph（ TSHi PSOG） method based on the framework of Hierarchical Probabilistic Self-Organizing Graph（ Hi PSOG） method in the single domain was proposed. Firstly,representation vectors with different granularities were extracted from source and target domains by using hierarchical self-organizing model based on a probabilistic mixture of multivariate Gaussian component; and the sparse graph probabilistic criterion was used to control the growth of the model. Secondly,the most similar representation vector of the target domain data was searched in the rich-information source domain by using the Maximum Information Coefficient（ MIC）. Then,the data in the target domain was classified using labels of similar representation vectors in the source domain. Finally,the experimental results on the international universal 20 Newsgroups dataset and the spam detection dataset show that the proposed method improves the average classifying accuracy of target domain using the information from source domain by 15. 26% and 9. 05%. Moreover,the approach improves the average classifying accuracy with mining different granularity representation vectors by 4. 48% and 4. 13%.