Learning process system for semi-supervised classification data from discriminatory parameters

摘要

Procedure semi- supervised learning computer for classification data (Di) executed, at least some of which are labeled with a label selected from a plurality of classes (Ck) predetermined, not being labeled the other, the method comprising: - one step (E20) of construction of a local model in the course of which each of said data (Di) is represented by a prototype (Pj); - a step (E30) labeling each prototype (Pj) with a label (Ej) comprising for each class (Ck) a probability (Pj, k) of said prototype (Pj) belongs to that class (Ck); - a step (E40) of defining a graph generator (G), whose nodes consist of these prototypes (Pj); - a stage (R40) allocation of a weight (WJ1, j, 2) to each of the arcs (aj1, j, 2) of this graph, being estimated this weight by the density (dj1, j, 2) data around this arc; - a step (E50) of propagation of said tags (Eg) in the degree (G) until a state of convergence, the spread of a disease associated with a first prototype (PJ1) to a second prototype (Pj, 2) attached tag said first prototype (PJ1) by an arch (aj1, j, 2) consisting in modifying the probabilities of the label (eg, 2) of the first prototype; and - a step (E60) classification of said data (Di) according to label prototypes representing said data after said propagation (E50) being assigned class untagged data according to the rule of maximum posteriori.