Hybrid neural networks for dimension reduction and clustering of multidimensional data

Understanding the information and clusters hidden inside multidimensional data can be challenging and complicated. Dimension reduction is usually considered as the first step for data analysis and interpretation. The focus of this paper is on the improvement of data clustering performance of Self Organising Maps (SOM) by embedding Auto-Associative Neural Networks (AANN). SOM is known as a computational tool that carries out topology preservation from high-dimensional input space onto a low-dimensional grid such as two-dimensional (2D) map. It has been used to visualize and explore inherent clusters and properties of the data. In this paper, a structurally flexible combination of AANN and SOM is developed, applied and investigated on Iris Flowers and Italian Olive oils datasets. The results have shown that the combined technique of AANNSOM has led to improvement of data clustering performance. It has reduced quantization error by 93.1%, and topographic error by 35.2%, when compared to SOM alone.