Advanced search
Start date
Betweenand


GEEK: Grammatical Evolution for automatically Evolving Kernel functions

Full text
Author(s):
Sousa, Arua de M. ; Lorena, Ana C. ; Basgalupp, Marcio P. ; IEEE
Total Authors: 4
Document type: Journal article
Source: 2017 16TH IEEE INTERNATIONAL CONFERENCE ON TRUST, SECURITY AND PRIVACY IN COMPUTING AND COMMUNICATIONS / 11TH IEEE INTERNATIONAL CONFERENCE ON BIG DATA SCIENCE AND ENGINEERING / 14TH IEEE INTERNATIONAL CONFERENCE ON EMBEDDED SOFTWARE AND SYSTEMS; v. N/A, p. 8-pg., 2017-01-01.
Abstract

One of the key aspects in the successful use of kernel methods such as Support Vector Machines is the proper choice of the kernel function. While there are several well known kernel functions which can produce satisfactory results for various applications (e.g. RBF), they do not take into account specific characteristics of the data sets. Moreover, they have a set of parameters to be tuned. In this paper, we propose GEEK, a Grammatical Evolution approach for automatically Evolving Kernel functions. GEEK uses a grammar composed of simple mathematical operations extracted from known kernels and is also able to optimize some of their parameters. When combined through the Grammatical Evolution, these operations give rise to more complex kernel functions, adapted to each specific problem in a data-driven approach. The predictive results obtained by Support Vector Machines using the GEEK kernel functions were in general statistically similar to those of the standard RBF, Polynomial and Sigmoid kernel functions, which had their parameters optimized by a grid search method. Nonetheless, the GEEK kernels were able to handle more properly imbalanced classification problems, whilst the results of the standard kernel functions were biased towards the majority class. (AU)

FAPESP's process: 12/22608-8 - Use of data complexity measures in the support of supervised machine learning
Grantee:Ana Carolina Lorena
Support Opportunities: Research Grants - Young Investigators Grants
FAPESP's process: 16/02870-0 - Multi-objective hyper-heuristics for automatic design of multi-test decision tree induction algorithms
Grantee:Márcio Porto Basgalupp
Support Opportunities: Regular Research Grants