Possible and necessary labels in K-nn procedures to query partially labelled data

Vu-Linh Nguyen (1) , Sébastien Destercke (1, 2) , Marie-Hélène Masson (1, 2, 3)

1 Heudiasyc - Heuristique et Diagnostic des Systèmes Complexes [Compiègne]
2 Labex MS2T - Laboratoire d'Excellence "Maîtrise des Systèmes de Systèmes Technologiques"
3 UPJV - Université de Picardie Jules Verne

When learning from partially labelled data (i.e., given as a subset of possible labels containing the true one), an issue that naturally arise is to determine which data should be queried to improve the accuracy of predictions, or in other word to determine an order between the partial labels to be queried. An answer to that question is to query the data that would induce the highest number of ambiguous predictions. In the K-nn case, studied in this paper, this question is similar to determining possible and necessary winners in plurality voting. In this paper, we discuss this connection as well as its use in partial label querying.

Domaines: Autre [cs.OH]

Vu-Linh Nguyen, Sébastien Destercke, Marie-Hélène Masson. Possible and necessary labels in K-nn procedures to query partially labelled data. From Multiple Criteria Decision Aid to Preference Learning, Nov 2016, Paderborn, Germany. ⟨hal-01396230⟩