Artificial neural network mapping experiments - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Reports (Contract/Project Report) Year : 2003

Artificial neural network mapping experiments


This report deals with a set of "artificial neural network mapping experiments". Artificial Neural Networks (ANNs) are a useful class of models for information visualization and analysis. Our interest in ANNs may be based on the links that exist between multivariate data analysis and the ANNs approaches in the areas of clustering and cartography.The ANNs experimented: Adaptative Resonance Theory (ART1), Multi-Layer Perceptron (MLP), SOM, Neural Gas, Growing Neural Gas, and Multi-SOM platform. The evaluation measures used: on the one hand, the mean squared error (MSE) and Quality defined as global relative error (GRE) (in chapters 3 and 4), on the other hand, Intra-Inertia, Inter-Inertia, and Quality reformulated as Recall, Precision, F function (in chapters 6 and 7). These two set of measurements are oriented to evaluated: the representation capability of maps (in chapters 3 and 4), and the clustering quality of the unsupervised clustering neural models (in chapter 6 and 7).Overview of the chapters:Chapter 1 is an introductory overview on what seems for us to be the current issues in information visualization and analysis with artificial neural networks (ANNs), according to our own experience in the field. Chapter 2 describes the extension of a two-step approach founded on axial k-means (AKM) for cluster analysis and principal component analysis (PCA) for representing the clusters on a map, into a neural network platform for cluster analysis and cartographyIn chapter 3, two different cartography approaches are compared and evaluated: principal component analysis (PCA) and Multi-Layer Perceptron, a 3-Layer and then a 4-Layer Perceptron (MLP) in "self-association" mode.The chapter 4 is a continuation of the former: we apply on a Web dataset the same two-step method AKM - MLP using the MLP architecture with two hidden layers.Chapter 5 deals with Multi-SOM as information analysis model. The principles and functionalities are reviewed from this angle. As leitmotiv "what kind of knowledge the user obtains with these operations?" The last point is that of the indicators in this visual-based information analysis model.Chapter 6 proposes new measures for estimating the quality of cluster analysis derived from the information retrieval (Precision, Recall, and F function). Two experiments using Multi-SOM are presented applying these measures.Chapter 7 deals with unsupervised neural methods as Neural Gas (NG), Growing Neural Gas (GNG), and Self-Organizing Map (SOM). The SOM, NG, and GNG networks realizing a global analysis which are compared with the Multi-SOM viewpoint-oriented-analysis
Not file

Dates and versions

inria-00099738 , version 1 (26-09-2006)


  • HAL Id : inria-00099738 , version 1


Claire François, Martial Hoffmann, Jean-Charles Lamirel, Xavier Polanco, Shadi Al Shehabi. Artificial neural network mapping experiments. [Contract] A03-R-359 || francois03b, 2003, 86 p. ⟨inria-00099738⟩
182 View
0 Download


Gmail Facebook Twitter LinkedIn More