Effects of missing data on neural network identification of biological taxa: RBF network discrimination of phytoplankton from flow cytometry data

摘要

Missing parameters are a common problem when trying to make biological identifications. Different ways of estimating missing parameter values are examined and identification success compared with that achieved when no parameters were missing and by networks trained without inclusion of the particular missing parameter, using radial basis function networks trained to discriminate 19 phytoplankton species on the basis of 11 flow cytometric parameters. Omitting each parameter singly from both training and test data sets resulted in only slightly reduced overall identification success; using maximum likelihood estimates of missing parameters gave only slightly less overall success than did nets trained on data sets with that parameter missing.