This paper proposes an original supervised method for learning a structuring model from a set of elements described by a collection of relations (multi-view context). It uses the theory of pretopology (and the crucial pseudo-closure operator) that offers a powerful formalism leading to complex structuring models. The pseudo-closure operator being non-idempotent, we show that the underlying binary classification problem matches with the well known multi-instance learning framework. We propose a multi-instance learning algorithm based on the enumeration of the positive and negative bags of instances rather than the instances themselves. Finally a proof of concept is proposed for the whole methodology that performs the task of lexical- taxonomy reconstruction.