Dimensional reduction and data visualization using hybrid artificial neural networks.

Data with dimension higher than three is not possible to be visualized directly. Unfortunately in real world data, not only the dimension are often more than three, very often real world data contain temporal information that makes the data only useful and meaningful when they are interpreted in sequence. Dimensionality reduction and visualization techniques such as self-organizing map (SOM) are usually used to explore the underlying multidimensional data structure. However, SOM only preserves inter-neurons distances in the input space and not in the output space due to the rigid grid used in SOM. Visualization induced self organizing map (ViSOM) was proposed as an extension of SOM in order to preserve the output space topology. In this paper, the modified adaptive coordinates (AC) technique is proposed to improve the visualization of SOM without the need to increase the number of neurons as in ViSOM. With a better visualization map formed, a post-processing technique is incorporated into the algorithm to produce a hybrid that is capable to extract temporal information contained in the data. Empirical studies of the hybrid techniques yield promising topology preserved visualizations and data structure exploration for synthetic and benchmarking datasets