查看“︁多示例学习”︁的源代码

在[[机器学习]]中， '''多示例学习''' (MIL) 是由[[監督式學習]]演变而来的。相较于输入一系列被单独标注的示例，在多示例学习中，输入的是一系列被标注的“[[包]]”，每个“包”都包括许多示例。举一个二元分类的简单的例子，当包中的所有示例都是负例时，这个包会被标注为负包。另一方面，当包中至少含有一个正例时，这个包会被标注为正包。当收到一系列被标注的包时，机器试着去：（1）归纳出一个类别概念以便正确标注个别示例。（2）在归纳之外学习怎样去标注一个包。

就图像分类举一个例子：给出一个图像，我们想要根据图像的画面内容来确定它的目标类别。比如，当图像同时包括了“沙子”和“水”时，图像的目标类别可能是“海滩”。在多示例学习中，图像被描述成一个包：<math>X = \{X_1,..,X_N\}</math>, 其中每一个<math>X_i</math>均是从图像中相应第i个区域中提取出来的特征向量（我们称之为示例），N是图像被分割出的区域（示例）个数。当图像包同时包含“沙子”区域示例和“水”区域示例时，这个包会被标注成正例（“海滩”）。
<span class="cx-segment" data-segmentid="9"></span>

多示例学习这一名称最初是由{{Harvtxt|Dietterich|Lathrop|Lozano-Pérez|1997}}提出来的，但是类似更早的研究，有{{Harvtxt|Keeler|Rumelhart|Leow|1990}}的手写数字识别。 最近关于多示例学习的回顾文献包括了{{Harvtxt|Amores|2013}}，对于不同的范式，它提供了一个广泛的回顾和比较研究。 还有{{Harvtxt|Foulds|Frank|2010}}，对于文献中不同的范式所提出的不同假设，它提供了一个全面的回顾。

运用多示例学习的几个例子：
* 分子活性
* 钙调素结合蛋白结合位点的预测 <ref>{{cite journal|last1 = Minhas|first1 = Fayyaz|title = Multiple instance learning of Calmodulin binding sites,|journal = Bioinformatics|date = 2012|volume = 28|issue = 18|page = i416-i422|url = http://bioinformatics.oxfordjournals.org/content/28/18/i416.short|doi = 10.1093/bioinformatics/bts416|access-date = 2015-07-10|archive-date = 2015-03-04|archive-url = https://web.archive.org/web/20150304025752/http://bioinformatics.oxfordjournals.org/content/28/18/i416.short|dead-url = no}}</ref>
* 对于选择性剪接异构体的预测作用 {{Harvtxt|Li|Menon|et al.|2014}},{{Harvtxt|Eksi|Li|Menon|et al.|2013}}
* 图像分类{{Harvtxt|Maron|Ratan|1998}}
* 文本或文档分类
数不清的研究都在做着促使传统分类技术，诸如[[支持向量机]]或是[[提升方法]]，适应于多示例学习环境的工作。

== 参见 ==
* [[Multi-label classification]]

== 参考资料 ==
{{Reflist}}
* {{Citation|first1 = Thomas G.|last1 = Dietterich|first2 = Richard H.|last2 = Lathrop|first3 = Tomás|last3 = Lozano-Pérez|title = Solving the multiple instance problem with axis-parallel rectangles|journal = Artificial Intelligence|volume = 89|issue = 1–2|year = 1997|pages = 31–71|doi = 10.1016/S0004-3702(96)00034-3}}.

* {{Citation|first1 = Jaume|last1 = Amores|title = Multiple instance classification: Review, taxonomy and comparative study|journal = Artificial Intelligence|volume = 201|year = 2013|pages = 81–105|doi = 10.1016/j.artint.2013.06.003}}.

* {{Citation|first1 = James|last1 = Foulds|first2 = Eibe|last2 = Frank|title = A Review of Multi-Instance Learning Assumptions|journal = Knowledge Engineering Review|volume = 25|issue = 1|year = 2010|pages = 1–25|doi = 10.1017/S026988890999035X}}.

* {{Citation|first1 = James D.|last1 = Keeler|first2 = David E.|last2 = Rumelhart|first3 = Wee-Kheng|last3 = Leow|title = Integrated segmentation and recognition of hand-printed numerals|year = 1990|pages = 557–563|work = Proceedings of the 1990 Conference on Advances in Neural Information Processing Systems (NIPS 3)}}.

* {{Citation|first1 = H.D.|last1 = Li|first2 = R.|last2 = Menon|first3 = |last3 = et al.|title = The emerging era of genomic data integration for analyzing splice isoform function|year = 2014|id = pii S0168-9525(14)00085-7|journal = Trends in Genetics|doi = 10.1016/j.tig.2014.05.005|pmid = 24951248}}.

* {{Citation|first1 = R.|last1 = Eksi|first2 = H.D.|last2 = Li|first3 = R.|last3 = Menon|first4 = |last4 = et al.|title = Systematically differentiating functions for alternatively spliced isoforms through integrating RNA-seq data|year = 2013|pages = Nov;9(11):e1003314|journal = PLoS Comput Biol|doi = 10.1371/journal.pcbi.1003314|pmid = 24244129|pmc = 3820534}}.

* {{Citation|first1 = O.|last1 = Maron|first2 = A.L.|last2 = Ratan|title = Multiple-instance learning for natural scene classification|year = 1998|pages = 341–349|work = Proceedings of the Fifteenth International Conference on Machine Learning}}.

* {{Cite conference|last1 = Ray|first1 = Soumya|first2 = David|last2 = Page|title = Multiple instance regression|conference = ICML|year = 2001|url = http://pages.cs.wisc.edu/~sray/papers/mip.reg.icml01.pdf|access-date = 2015-07-10|archive-date = 2022-02-13|archive-url = https://web.archive.org/web/20220213063446/https://pages.cs.wisc.edu/~sray/papers/mip.reg.icml01.pdf}}.
[[Category:机器学习]]