Estimating Null Values in Relational Databases Using Analogical Proportions

This paper presents a novel approach to the prediction of null values in relational databases, based on the notion of analogical proportion. We show in particular how an algorithm initially proposed in a classification context can be adapted to this purpose. This work focuses on the case of a transactional database, where attributes are Boolean. The experimental results reported here, even though preliminary, are encouraging since the approach yields a better precision, on average, than the classical nearest neighbors technique.