Learning with Augmented Multi-Instance View

[edit]

Yue Zhu, Jianxin Wu, Yuan Jiang, Zhi-Hua Zhou ;
Proceedings of the Sixth Asian Conference on Machine Learning, PMLR 39:234-249, 2015.

Abstract

In this paper, we propose the Augmented Multi-Instance View (AMIV) framework to construct a better model by exploiting augmented information. For example, abstract screening tasks may be difficult because only abstract information is available, whereas the performance can be improved when the abstracts of references listed in the document can be exploited as augmented information. If each abstract is represented as an instance (i.e., a feature vector) x, then with the augmented information, it can be represented as an instance-bag pair (x;B), where B is a bag of instances (i.e., the abstracts of references). Note that if x has a label y, then we assume that there must exist at least one instance in the bag B having the label y. We regard x and B as two views, i.e., a single-instance view augmented with a multi-instance view, and propose the AMIV-lss approach by establishing a latent semantic subspace between the two views. The AMIV framework can be applied when the augmented information is presented as multi-instance bags and to the best of our knowledge, such a learning with augmented multi-instance view problem has not been touched before. Experimental results on twelve TechPaper datasets, five PubMed data sets and a WebPage data set validate the effectiveness of our AMIV-lss approach.

Related Material