Background: A general restrictions-free theory supporting a functional dependency between the data, the artificial classification algorithm's internal specifics and its performance (e.g. error) has not yet been devised. Thus, in the design phase of building a classification tree, an important choice of selecting an appropriate algorithm family must be made.
Objective: The objective of this paper is to compare the Support Vector Machines and of the Neural Network classification algorithm on real data sets in terms of accuracy, and in terms of a function that best describes the error rate.
Methods: A Weka-based multilayer perceptron (MLP) neural network and a Support Vector Machines classification algorithms were applied to a set of datasets (n=121) from publicly available repositories (UCI) in step wise k-fold cross-validation and an error rate was measured in each step. First, four different functions, i.e. power, linear, logarithmic, exponential, were fit to the measured error rates. Where the fit was statistically significant (n=54) for all functions for both algorithms, we measured the average mean squared error rate for each function and its rank. The Wilcoxon's signed rank test was used to test whether the differences between ranks are significant.
Results: In a total of 54 datasets, the SVM algorithm using the exponential function was performing better than NN (P=0,023). Average mean squared error using all datasets and the exponential function for description of learning statistically significantly differ from each other. The exponential function is thus best describing the learning process. The chosen neural network and support vector machines algorithms are not different from each other in the capacity of capturing the interrelations in the data. However, SVM gives better results when using the exponential function.
Conclusion: The exponential model can be used to forecast the future performance of both algorithms based on a small training sample. The selection process of the two algorithms shows that both algorithms are equal on capturing the data interrelations, but support vector machines is yielding lower error rates.