Adaptive Kernel Data Streams Clustering Based on Neural Networks Ensembles in Conditions of Uncertainty About Amount and Shapes of Clusters

Abstract

The neural network's approach for data stream clustering task, that in online mode are fed to processing in assumption of uncertainty about amount and shapes of clusters, is proposed in the paper. The main idea of this approach is based on the kernel clustering and idea of neural networks ensembles, that consist of the T. Kohonen's selforganizing maps. Each of the clustering neural networks consists of different number of neurons, where number of clusters is connected with the quality of these neurons. All ensemble members process information that sequentially is fed to the system in the parallel mode. Experimental results have proven the fact that the system under consideration could be used to solve a wide range of Data Mining tasks when data sets are processed in an online mode.

Online Fuzzy Clustering of High Dimension Data Streams Based on Neural Network Ensembles

“…The last expression of (11) for calculating centroids of clusters can be rewritten in recurrent form. which essentially coincides with the self-learning WTMrule by T. Kohonen [5], where the factor Thus, the process of high-dimensional data clustering (11), (12) is conveniently implemented using the architecture shown in fig. 4 which is a modification of the neuro-fuzzy network of T. Kohonen [26,27].…”

“…4 which is a modification of the neuro-fuzzy network of T. Kohonen [26,27]. Here, the first hidden kernel layer (KL) is essentially a standard SOM neural network [5], which contains mneurons in the Kohonen layer, whose synaptic weightscentroids are tuned using the WTM learning rule (12), in the second hidden layer ML, the membership levels of kth observation to j-th cluster   j uk using the first relation (11) are estemated, and in the output layer WL weights ji  are calculated using the second relation of (11).…”

“…Thus, if data sample is a set (possibly growing) , which is fed to the inputs of an ensemble formed by 1 M  parallelconnected SOMs so that the first one works in conditions that the number of possible clusters in the data is 2 m  , and the last assumes that mM  , the best results will be obtained using SOM with * 2 mM  neurons in the Kohonen layer, where * m determines the true number of classes in the sample under processing [12].…”

“…the number of neurons in the Kohonen`s layer KL, and p  is the index of a specific fuzzifier, taking q values. All elements are tuned using the same type of procedures(11),(12), which differ from each other only in the values of m and  .In blocks[] m pMEXB , the quality of clustering provided by a particular FSOM is evaluated, and the output layer of the DM ensemble selects the best from   Mq  results of the previous layers, i.e. the number of clusters m* in the processed data, the centroids of the formed clusters…”

Ансамбль Нейромереж GRNN На Підставі Зміщених Поверхонь Відгуку Для Задач Електронної Комерції

“…Штучні нейронні мережі успішно зарекомендували себе у розв'язанні цієї задачі. Вони автоматично будують високоточні моделі, які можуть аналізувати великі та складноструктуровані дані, але незважаючи на велику кількість успішних застосувань, ці методи мають низку недоліків (Bodyanskiy, Tyshchenko & Kopaliani, 2018;Zhernova et al, 2018).…”