A multi-armed bandit approach for batch mode active learning on information networks

We propose an adaptive batch mode active learning algorithm, MABAL (Multi-Armed Bandit for Active Learning), for classification on heterogeneous information networks. Observing the parallels between active learning and multi-armed bandit (MAB), we base MABAL on an existing combinatorial MAB algorithm to combine simple strategies to generate query batches. MABAL employs a novel error expectation measure for network classification that does not assume assortativity as MAB reward feedback to determine the most fit strategy for the given task. We provide a preliminary optimality analysis of MABAL based on performance bounds for combinatorial MAB. A case study illustrates that MABAL not only converges quickly to the optimal strategy but also provides insight into the functional roles of the different node types. Evaluations of MABAL on real world network classification tasks demonstrate that it achieves performance gains over existing methods independent of the underlying classification model.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2016

Author dc:creator

• Liao, De

Contributors dc:contributor

• Han, Jiawei

• Active learning
• Heterogeneous information networks
• Multi-armed bandit

Statement dc:rights

• Copyright 2016 De Liao

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/90788

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/90788