Sequential Feature Selection and Instance Selection Using SpFSR and SpFixedIS

In this paper two-dimensional data reduction is examined in a sequential manner for the emerging class of optimisation-based wrapper methods, namely SpFixedIS for instance selection and SpFSR for feature selection. The study examines the performance of instance selection on a feature reduced dataset (IFS) and the performance of feature selection on an instance reduced dataset (FIS). The results show that on average FIS marginally outperforms IFS across all 6 of the different classifiers, however the majority of tests are statistically equivalent. IFS provides less variability in performance compared to FIS regardless of selection rate. However, the performance of the one-dimensional reductions generally outperformed their two-dimensional counterparts. Whilst utilising the Gaussian Naive Bayes wrapper, FIS and IFS provide an average improvement over the full training set of 9.281% and 8.170% respectively. This study is motivated by future work to propose novel optimisation-based wrapper hybrid selection algorithms, that is data reduction approaches which can perform simultaneous instance and feature selection.