Ubiquitous Self-Organizing Maps

Abstract

Knowledge discovery in ubiquitous environments are usually conditioned by the data stream model, e.g., data is potentially infinite, arrives continuously and is subject to concept drift. These factors present additional challenges to standard data mining algorithms. Artificial Neural Networks (ANN) models are still poorly explored in these settings. State-of-the-art methods to deal with data streams are single-pass modifications of standard algorithms, e.g., K-means for clustering, and involve some relaxation of the quality of the results, i.e., since the data cannot be revisited to refine the models, the goal is to achieve good approximations [Gama, 2010]. In [Guha et al., 2003] an improved single pass k-means algorithm is proposed. However, k-means suffers from the problem that the initial k clusters have to be set either randomly or through other methods. This has a strong impact