Slade Matthews, Ian Spence, Herbert F. Jelinek, and Craig S. McLachlan
Machine learning algorithms, balanced datasets, verification, performance, neurons
Retinal ganglion cells of different species have been categorised using different paradigms and resulting in different number of suggested classes/types based on traditional morphological parameters such as cell body size or dendritic diameter. The inherent nature of the neuron’s branching pattern has also been shown to play a role in signal processing and therefore additional features such as fractal dimension and lacunarity were added. Machine learning algorithms (MLA) provide a basis for classification tasks based on large numbers of features. However there are numerous ways of presenting data, different algorithms, validation methods and determination of performance. No studies have been undertaken that investigate the influence of model validation on small datasets with diverse feature parameters. This paper outlines the differences when balanced and imbalanced data is used in combination with six supervised MLAs and two different validation algorithms (LOO and 10-fold) as well as interpreting the results using two performance measures (AUC or accuracy). Our results indicate that that the largest effect on MLA outcomes is whether data is balanced or imbalanced. AUC is a more robust decision rule compared to accuracy. The best classifiers for our data were neural networks and logistic regression with an AUC of greater than 0.9.
Important Links:
Go Back
