Functionally expanded streaming data as input to classification processes using ensembles of constructive neural networks
João Roberto Bertini, Maria do Carmo Nicoletti · 2016
Machine learning (ML) based applications that require data stream processing have become quite common over the past few years. To deal with continuous and massive streams of data, low computational and memory costs are required from the ML techniques employed; these requirements can be partially fulfilled by using constructive neural networks (CoNN) algorithms. The automatic definition of the Neural Network (NN) architecture, as well as its fast training, promote the high adaptability of such NNs, which allows to skip the conventional model selection phase during the learning process. However, such advantages usually come with the drawback of promoting lower accuracy rates, when compared to other learning approaches. The work described in this paper proposes an ensemble of CoNNs combined with functional expansion techniques to cope with data stream classification. The experiment results, followed by a comparative analysis, showed that CoNN accuracy performance along streaming data does improve with the use of functional expansion. Considering that to functionally expand the input data has a low computational cost, the obtained results turn CoNN algorithms an interesting approach to be considered in data stream classification.