Cooling Growing Grid: An Incremental Self-organizing Neural Network For Data Exploration
Free (open access)
V. Tomenko & V. Popov
Fundamental self-organizing artificial neural networks, both static (with predefined number of neurons) and incremental, are presented and goals of competitive learning are enumerated. A novel incremental self-organizing ANN Cooling Growing Grid (CGG) is proposed, which combines the advantages of static and incremental approaches and overcomes their main drawbacks. The estimation of growth direction is made less haphazard and the weights adaptation rule is modified in order to achieve better performance for highly nonuniform real-world data. The main performance measures for evaluation of reference vectors distribution are utilized in order to compare CGG with existing models. Keywords: self-organizing neural network, incremental, data exploration. 1 Introduction Exploratory data analysis is the data-driven search for statistical insights and models, resulting in understandable data set representation and preservation of as much essential information in the data, as possible. Exploratory data analysis is used as a tool in data mining. Self-organizing ANNs are extremely important and useful for exploratory data analysis. Their main distinctive feature is ability to \“represent” similar in some sense input samples (real-valued vectors) by single prototype vector. Therefore, such networks are able to extract significant features from data sets. Furthermore, the results can be visualized in order to find groups (clusters, categories) of similar patters and understand their mutual arrangement. On the other hand, self-organizing networks together with auxiliary techniques can be used to control robots, for data compression, prediction. Such wide applicability resulted in development of different variants of self-organizing
self-organizing neural network, incremental, data exploration.