Comparison of artificial neurol networks classification ability with factor analysis and artificial neurol network hybrid model

Aim: Compared to the Classical Artifical Neurol Network (Classical ANN) model created with estimated parameters, by taking advantage of the ability of Factor Analysis to define the association in variables according to causality, by developing a Hybrid Artifical Neurol Network model that can create in a very short period of time the ANN architecture similar to the data topology determined by the factor analysis, it is aimed to classify the data more successfully and in a much shorter period of time compared to the Classical Artifical Neurol Network. Methods: The codes of the Classical Artificial Neural Networks classifier method have been modified on Jupyter Notebook platform. In the algorithm of the Hybrid ANN model, in the first step, factor analysis is applied to the data. In the second step, codes were written to design the architecture of the Artificial Neural Network model in a way that it has as many layers as the number of factors that explain the 0.70 variance of the data and as many neurons as the number of variables. In the case of single-factor data, the variables of factors which loads less than 0.20 is removed from the data so that it has as many neurons as the number of remaining variables and a single layer. In the third step, Hybrid ANN with the determined architecture assigns the data to classes. Results: Five data sets with normal distribution, consisting of continuous numerical values (x∼N(μ,σ^2), n=5,000), with five, four, three, two, and one factors, and 20, 16, 12, 8 and 10 variables, respectively, have been created. The accuracy scores of the most successful Classical ANN model and the Hybrid ANN model are over 0.90. While creating the architecture and training of the Classical ANN models take 240 minutes, the Hybrid ANN model creates its own architecture and completes its training in 0.2 seconds. Conclusions: The Hybrid ANN is a successful ANN model determinant alternative. Keywords: Factor Analysis, Hybrid Neural Networks, Data Generation, Python