Multi-Label Based Learning for Better Multi-Criteria Ranking of Ontology Reasoners Nourh` ene Alaya 1,2 , Myriam Lamolle 1 , and Sadok Ben Yahia 2 LIASD, University of Paris 8, IUT of Montreuil, Saint-Denis, France. 1 LIPAH, University of Tunis El-Manar, Faculty of Sciences, Tunis, Tunisia. 2 {n.alaya,m.lamolle}@iut.univ-paris8.fr, sadok.benyahia@fst.rnu.tn Abstract. A growing number of highly optimized reasoning algorithms have been developed to allow inference tasks on expressive ontology lan- guages such as OWL(DL). Nevertheless, there is broad agreement that a reasoner could be optimized for some, but not all the ontologies. This particular fact makes it hard to select the best performing reasoner to handle a given ontology, especially for novice users. In this paper, we present a novel method to support the selection ontology reasoners. Our method generates a recommendation in the form of reasoner ranking. The efficiency as well as the correctness are our main ranking criteria. Our solution combines and adjusts multi-label classification and multi-target regression techniques. A large collection of ontologies and 10 well-known reasoners are studied. The experimental results show that the proposed method performs significantly better than several state-of-the-art rank- ing solutions. Furthermore, it proves that our introduced ranking method could effectively be evolved to a competitive meta-reasoner. Keywords: Ontology, Reasoner, Multi-label classification, Multi-target regression, Multi-criteria, Ranking, Advising, Meta-reasoning 1 Introduction A growing number of highly optimized ontology reasoners [10] have been devel- oped to allow inference tasks on expressive ontology languages such as OWL(DL) [6]. Nevertheless, it is well accepted that a reasoner could be optimized for some but not all the ontologies. Indeed, the respective authors of [18, 5] have outlined that, often in practice, reasoners tend to exhibit unpredictable behaviours when dealing with real world ontologies. They noticed that the reasoner performances can considerably vary across the ontologies, even when the size or/and the ex- pressivity of these ones are fixed. Furthermore, Gardiner et al. [4] and more recently Lee et al. [9] pinpointed out that reasoners may disagree over inferences or query answers, computed from the same input ontology. All of the aforemen- tioned authors offered different explanations of these phenomena: bottlenecks in the ontology design [5]; interactions between reasoning optimisation techniques [4]; or even reasoner implementation bugs [9]. Given all of these findings, it is obvious that for a typical OWL user, deciding the most performing reasoner to handle a given ontology is not a trivial task.