摘要
针对现有多标签数据集转换方法无法有效利用标签间的语义相关性和共现性知识,以及转换得到的数据集相对于问题规模偏小等问题,提出了一种新的多标签数据集转换方法 RAPC-W(ranking by all pairwise com-parision based WordNet)。该方法将标签对从原来的两对扩展到四对,增加了划分后数据集的规模。另外,引入了外部数据源WordNet,较好地考虑了标签语义相关性和共现性知识,一定程度上过滤掉了语义不相关的标签组合,更好地保留了原始数据集的信息,降低了噪声数据集对基分类器训练的不良影响。在UCI知识库提供的Yeast和Letter数据集以及KEEL提供的Emotion、Genbase数据集上的一系列实验结果表明,该方法是有效可行的。
Existing multi-label data set transformation method can not effectively utilize semantic correlation between the label and the co-occurrence of knowledge, as well as the dataset is small-scale to the scale of the problem such, this paper presented a new multi-label data set into method RAPC-W（ ranking by all pairwise eomparision based WordNet） , the method extend the label pair from the original two pairs to four pairs increasing the size of the devided data set. On the other hand introduced external data sources WordNet, taking a comprehensive consideration of the label semantic correlation and the co-occurrence of knowledge, to some extent, filter out the uncorrelated label combination in semantics, better to retain the information of the original data set, also reduce the adverse effects of the noise data set to the based classifier training. A series of experimental results based on the Yeast and Letter data set provided by the UCI knowledge as well as the Emotion and Genhase data set provided by the KELL shows that this method is effective and feasible.
出处
《计算机应用研究》
CSCD
北大核心
2013年第6期1692-1695,共4页
Application Research of Computers
基金
国家自然科学基金资助项目(71271117)
江苏省科技型企业技术创新资金资助项目(BC2012331)
