Given a set of categorical variables, we want to predict one or more of them by the way rules. We propose an algorithm that (i) is guided by statistical results in a relational geometry where we use assymetrical association indices, and (ii) makes statistical and euclidian approximations. The iterative method we propose can obtain rules without introducing a priori their premises in the set of independent conjonctions analized by the generator at each step. The algorithm has a linear complexity with regard to the number of individual; this property makes it suitable for large data sets. We present results over data examples.
