From the products bought, by basket analysis we seek to infer interest, values and choice criteria and predict purchase probabilities for other products. This statistical approach relies on the existence of a few general under-lying clusters which enables the prediction of general and specific buying behavior. Compared to traditional clustering methods, a Kohonen map, a neural network, allows the projection and clustering of data for which the proximity presents a meaning or interest. Beyond the interest of these neural networks for graphic representation, this article suggests different ways of articulating general and product-specific typologies which are illustrated on a real database of buyers' behavior in a book club. The results clearly show a significant improvement with regards to the results obtained with current models using either RFM segmentation or logistic regression.