Artificial neural networks constitute good tools for certain types of computational modeling (being potentially efficient, easy to adapt and fast). However, they are often considered difficult to interpret, and are sometimes treated as black boxes. However, whilst this complexity implies that it is difficult to understand the internal organization that develops through learning, it usually encapsulates one of the key factors for obtaining good results. First, to yield a better understanding of how artificial neural networks behave and to validate their use as knowledge discovery tools, we have examined various theoretical works in order to demonstrate the common principles underlying both certain classical artificial neural network, and statistical methods for regression and data analysis. Second, in light of these studies, we have explained the specificities of some more complex artificial neural networks, such as dynamical and modular networks, in order to exploit their respective advantages in constructing a revised model for knowledge extraction, adjusted to the complexity of the phenomena we want to model. The artificial neural networks we have combined (and the subsequent model we developed) can, starting from task data, enhance the understanding of the phenomena modeled through analyzing and organizing the information for the task. We demonstrate this in a practical prediction task for telecommunication, where the general domain knowledge alone is insufficient to model the phenomena satisfactorily. This leads us to conclude that the possibility for practical application of our work is broad, and that our methods can combine with those already existing in the data mining and the cognitive sciences.