Scribd
Upload
18 views1 page

Active Learning With Multiple Views

Active Learning with Multiple Views

Uploaded by

Srinivas
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
18 views1 page

Active Learning With Multiple Views

Active Learning with Multiple Views

Uploaded by

Srinivas
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

 Section: Unlabeled Data

Active Learning with Multiple Views


Ion Muslea
SRI International, USA

INTRODUCTION shown by Nigam, et al. (2000) and Raskutti, et al. (2002),


their distribution can be used to boost the accuracy of a
Inductive learning algorithms typically use a set of classifier learned from the few labeled examples.
labeled examples to learn class descriptions for a set of Intuitively, semi-supervised, multi-view algorithms
user-specified concepts of interest. In practice, labeling proceed as follows: first, they use the small labeled
the training examples is a tedious, time consuming, er- training set to learn one classifier in each view; then,
ror-prone process. Furthermore, in some applications, they bootstrap the views from each other by augmenting
the labeling of each example also may be extremely the training set with unlabeled examples on which the
expensive (e.g., it may require running costly labora- other views make high-confidence predictions. Such
tory tests). In order to reduce the number of labeled algorithms improve the classifiers learned from labeled
examples that are required for learning the concepts data by also exploiting the implicit’ information pro-
of interest, researchers proposed a variety of methods, vided by the distribution of the unlabeled examples.
such as active learning, semi-supervised learning, and In contrast to semi-supervised learning, active
meta-learning. learners (Tong & Koller, 2001) typically detect and ask
This article presents recent advances in reducing the user to label only the most informative examples
the need for labeled data in multi-view learning tasks; in the domain, thus reducing the user’s data-labeling
that is, in domains in which there are several disjoint burden. Note that active and semi-supervised learners
subsets of features (views), each of which is sufficient take different approaches to reducing the need for la-
to learn the target concepts. For instance, as described beled data; the former explicitly search for a minimal
in Blum and Mitchell (1998), one can classify segments set of labeled examples from which to perfectly learn
of televised broadcast based either on the video or on the target concept, while the latter aim to improve a
the audio information; or one can classify Web pages classifier learned from a (small) set of labeled examples
based on the words that appear either in the pages or by exploiting some additional unlabeled data.
in the hyperlinks pointing to them. In summary, this In keeping with the active learning approach, this
article focuses on using multiple views for active learn- article focuses on minimizing the amount of labeled data
ing and improving multi-view active learners by using without sacrificing the accuracy of the learned classi-
semi-supervised- and meta-learning. fiers. We begin by analyzing co-testing (Muslea, 2002),
which is a novel approach to active learning. Co-testing
is a multi-view active learner that maximizes the benefits
BACKGROUND of labeled training data by providing a principled way
to detect the most informative examples in a domain,
Active, Semi-Supervised, and thus allowing the user to label only these.
Multi-view Learning Then, we discuss two extensions of co-testing
that cope with its main limitations—the inability to
Most of the research on multi-view learning focuses on exploit the unlabeled examples that were not queried
semi-supervised learning techniques (Collins & Singer, and the lack of a criterion for deciding whether a task
1999, Pierce & Cardie, 2001) (i.e., learning concepts from is appropriate for multi-view learning. To address the
a few labeled and many unlabeled examples). By them- former, we present Co-EMT (Muslea et al., 2002a),
selves, the unlabeled examples do not provide any direct which interleaves co-testing with a semi-supervised,
information about the concepts to be learned. However, as multi-view learner. This hybrid algorithm combines
the benefits of active and semi-supervised learning by

Copyright © 2009, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited.

You might also like