Mining association rules from clustering models

Association rules used to discover the interested relationships between variables (items) in large database. These relationships have been hidden in large database. Symbolic models consider the most common to extract association rules. Unfortunately, these models suffered of very serious limitations. Association rule generation is a process that take long time to generate a huge number of rules, including large amount of redundant rules. This problem appears obviously specially when dealing with high dimensional description space. To cope with these problems unsupervised approach has been made. This approach has been proposed to extract numerical association rules by establishing interesting links between numerical and symbolic models. Numerical models keep only the important relations between data, so it can extract the most potential and important association rules. This numerical approach can extract only simple association rules that mean each rule has two items, one item in each part of the rule. So it generates a lot of simple rules. Therefore, the aim of this thesis is to develop this approach to be able to extract complex and important association rules in order to discover relationships between itemsets of different sizes instead of only two items. This approach increases the possibility of numeric methods to extract important association rules. Neural network model with multiple topographic considers the main model in this thesis. One of the strongest features from this model is generalization mechanism that allows association rule extraction from only one database to be performed. The extraction of association rules is itself based on quality measures which evaluate to what extent a numerical classification model behaves as a natural symbolic classifier.