Learning From Positive and Unlabeled examples

Fabien Letouzey (1) , François Denis (1) , Rémi Gilleron (1)

1 GRAppA - LIFL - Groupe de Recherche en Apprentissage Automatique

In many machine learning settings, examples of one class (called positive class) are easily available. Also, unlabeled data are abundant. We investigate in this paper the design of learning algorithms from positive and unlabeled data only. Many machine learning and data min ing algorithms use examples for estimate of probabilities. Therefore, we design an algorithm which is based on positive statistical queries (estimates for probabilities over the set of positive instances) and instance statistical queries (estimates for probabilities over the instance space). Our algorithm guesses the weight of the target concept (the ratio of positive instances in the instance space) with the help of a hypothesis testing algorithm. It is proved that any class learnable in the Statistical Query model [Kea93] such that a lower bound on the weight ofany target concept f can be estimated in polynomial time is learnable from positive statistical queries and instance statistical queries only. Then, we design a decision tree induction algorithm POSC4.5, based on C4.5 [Qui93], using only positive and unlabeled examples. We alsogive experimental results for this algorithm.

Domaines: Langage de programmation [cs.PL]

Fabien Letouzey, François Denis, Rémi Gilleron. Learning From Positive and Unlabeled examples. Proceedings of the 11th International Conference on Algorithmic Learning Theory, ALT'00, 2000, Sydney, Australia. pp.71--85. ⟨inria-00538887⟩