Coevolution of mapping functions for linear SVM

Created by W.Langdon from gp-bibliography.bib Revision:1.4208

  author =       "Satish Kumar Jaiswal and Hitoshi Iba",
  booktitle =    "2017 IEEE Congress on Evolutionary Computation (CEC)",
  title =        "Coevolution of mapping functions for linear SVM",
  year =         "2017",
  editor =       "Jose A. Lozano",
  pages =        "2225--2232",
  address =      "Donostia, San Sebastian, Spain",
  publisher =    "IEEE",
  isbn13 =       "978-1-5090-4601-0",
  abstract =     "A linear SVM scales linearly with the size of a
                 dataset, and hence is very desirable as a classifier
                 for large datasets. However, it is not able to classify
                 a dataset having a nonlinear decision boundary between
                 the classes unless the dataset has been transformed by
                 some mapping function so that the decision boundary
                 becomes linear or it is a good approximation to a
                 linear boundary. Often these mapping functions may
                 result in a dataset with very large dimension or even
                 infinite dimension. To avoid the curse of
                 dimensionality, kernel functions are used as mapping
                 functions. However, a kernel SVM has quadratic time
                 complexity, and hence does not scale very well with
                 large datasets. Moreover, the choice of a kernel
                 function and its parameter optimization are arduous
                 tasks. Therefore, a replacement of kernel function with
                 an explicit mapping function is desirable in the case
                 of large datasets. In this paper, we propose a novel
                 co-evolutionary approach to find an explicit mapping
                 function. We use GA to evolve an n-tuple of GP trees as
                 a mapping function, and GP to evolve each individual GP
                 tree. The dataset is then transformed using the found
                 mapping function so that a linear SVM can be used.
                 Besides the fact that the proposed algorithm allows us
                 to use a fast linear SVM, the results also show that
                 the proposed algorithm outperforms the kernel trick and
                 even performs as good as the kernel trick combined with
                 feature selection.",
  keywords =     "genetic algorithms, genetic programming, computational
                 complexity, feature selection, pattern classification,
                 support vector machines, trees (mathematics), GA, GP
                 tree n-tuple evolution, coevolutionary approach,
                 dataset classifier, explicit mapping function, infinite
                 dimension, kernel functions, linear SVM, mapping
                 function coevolution, nonlinear decision boundary,
                 parameter optimization, quadratic time complexity,
                 Kernel, Optimization, Sociology, Statistics, Symbiosis,
                 Vegetation, co-evolutionary algorithm, feature
                 extraction, feature map, genetic algorithm, mapping
  isbn13 =       "978-1-5090-4601-0",
  DOI =          "doi:10.1109/CEC.2017.7969574",
  month =        "5-8 " # jun,
  notes =        "IEEE Catalog Number: CFP17ICE-ART Also known as

Genetic Programming entries for Satish Kumar Jaiswal Hitoshi Iba