For visualization-based analysis tools in knowledge discovery process: A multilayer perceptron versus principal components analysis: A comparative study

Mapping knowledge structures is a key task in Knowledge Discovery in Databases (KDD). In order to display the thematic organization of knowledge, we compare and evaluate two different cartography approaches: principal components analysis (PCA) and a multilayer perceptron (MLP) in “self-association” mode. This kind of MLP can be used to perform a PCA when the activation function is set to the identity function. This allows us to look for the non-linear activation function which best fits the data structure. We present an evaluation criterion and the results and maps obtained with both methods. We notice that the MLP detects a non-linearity in the data structure that the PCA does not detect. However, the MLP does not express the non-linearity completely. Finally we show how a related component analysis (RCA), based on graph theory, provides representations of the inter-clusters relationships, compensating for the approximate nature of the maps, and improving their readability.