Skip to Main content Skip to Navigation

De l'induction confirmatoire à la classification : contribution à l'apprentissage automatique

Abstract : Confirmatory generalisation consists in determining the most general laws confirmed by a set of observations. Confirmatory induction is based on the similarity assumption: unknown individuals behave like known individuals. We show that a circumscription of individuals is more appropriate to model this assumption than a circumscription of properties. Compared to existing approaches, the model we propose can produce more general clauses and avoids the production of unwanted generalisations. We specialise this model defined in first-order logic to attribute-value languages and show that it cornes down, in this case, to the calculation of prime implicates. Considering the problem of classification of objects from examples, we show that the minimal consistent and relevant rules are not, in general, confirmatory generalisations. We propose a new classification technique, called scope classification, which consists of building the set of examples from which a consistent and relevant rule can be built, so it is an instance-based approach of rule-based classification. We present several adaptations of the logical basements of the scope classification to better deal with real data. Scope classification is also extended to instances generalised into rules. The generalisation strategies we introduce, and especially the search of the neighbours, clearly differ from those of existing techniques. We show that, whereas hypotheses built by the scope classification and by the disjunctive version space differ, both techniques leads to the same classification . Though, in addition to a more efficient implementation, our original point of view allows us to propose sorne developments specifie to a rule-based approach. Our system obtains on average a better accuracy and a similar execution time to sorne of the most used instance-based or rule-based systems on an usual set of benchmarks.
Document type :
Complete list of metadata
Contributor : Thèses UL Connect in order to contact the contributor
Submitted on : Thursday, March 29, 2018 - 11:02:34 AM
Last modification on : Saturday, October 16, 2021 - 11:26:08 AM
Long-term archiving on: : Thursday, September 13, 2018 - 6:44:45 PM


Files produced by the author(s)


  • HAL Id : tel-01747518, version 1



Nicolas Lachiche. De l'induction confirmatoire à la classification : contribution à l'apprentissage automatique. Autre [cs.OH]. Université Henri Poincaré - Nancy 1, 1997. Français. ⟨NNT : 1997NAN10267⟩. ⟨tel-01747518⟩



Record views


Files downloads