Multi-Objective Genetic Programming for Dataset Similarity Induction

0455776 - ÚI 2016 RIV US eng C - Conference Paper (international conference)
Šmíd, J. - Pilát, M. - Pešková, K. - Neruda, Roman
Multi-Objective Genetic Programming for Dataset Similarity Induction.
SSCI 2015 IEEE Symposium Series on Computational Intelligence. Los Alamitos: IEEE, 2015, s. 1576-1582. ISBN 978-1-4799-7560-0.
[SSCI 2015. Symposium Series on Computational Intelligence. Cape Town (ZA), 08.12.2015-10.12.2015]
R&D Projects: GA ČR GA15-19877S
Grant - others:GA UK(CZ) 610214
Institutional support: RVO:67985807
Keywords : meta-learning * metric induction * multi-objective optimization * genetic programming
Subject RIV: IN - Informatics, Computer Science

Metalearning - the recommendation of a suitable machine learning technique for a given dataset - relies on the concept of similarity between datasets. Traditionally, similarity measures have been constructed manually, and thus could not precisely grasp the complex relationship among the different features of the datasets. Recently, we have used an attribute alignment technique combined with genetic programming to obtain more fine-grained and trainable dataset similarity measure. In this paper, we propose an approach based on multiobjective genetic programming for evolving an attribute similarity function. Multi-objective optimization is used to encourage some of the metric properties, thus contributing to the generalization abilities of the similarity function being evolved. Experiments are performed on the data extracted from the OpenML repository and their results are compared to a baseline algorithm.
Permanent Link: http://hdl.handle.net/11104/0256400