Automated online feature selection and learning from high-dimensional streaming data using an ensemble of Kohonen neurons

Here we describe a new algorithm that uses Kohonen networks at its core for class-based feature selection and for learning to recognize classes of patterns. This online algorithm is meant for streaming big data and for highly parallel implementation on a platform such as Apache Spark. The algorithm works in two phases. In the initial phase, it examines some streaming data to determine the features that distinguish a particular class from the rest of the classes. After this phase of class-based feature selection, it then uses those selected features to learn pattern classifiers. All phases use Kohonen networks and Kohonen style online learning. Kohonen networks trained in the first phase are discarded once features are selected. Automation is based on an ensemble of Kohonen neurons for pattern classification. We provide here some initial computational results on some high-dimensional gene expression problems based on a desktop implementation. In testing this algorithm, no parameters were changed for the different problems solved. And that is an essential feature of automation of learning.