The paper considers the issue of effective formation of a representative sample to train a neural network of a multilayer perceptron (MLP). As it is known, the key problem of MLP training is the factor space division into the test, validation and training sets. To solve this problem, an approach based on the use of clustering and a Lipschitz constant estimate is put forward. Kohonen's self-organizing maps (SOM) prove to be an effective clustering method. Based on such maps, the factor spaces of different dimensions are clustered and a representative sample is formed. Special software on C# programming language to calculate Lipschitz constant is developed. It allows Lipschitz constant calculation and subsequent clarification of factor space clusterization. To evaluate the effectiveness of the suggested approach, two training procedures are applied to the MLP neural network — with and without the use of clusterization with an estimate of Lipschitz constant. It is concluded that the approach under consideration influences the increase of the entropy as well as the decrease of the mean square error of the training set and, as a result, leads to the quality improvement of MLP neural network training with the small dimension of the factor space. Also, based on Lipschitz constant estimation, it is concluded that the training set entropy is not the only factor that has the main impact on the MLP training result.