Skip to Main content Skip to Navigation

Neural learning and validation of hierarchical multi-criteria decision aiding models with interacting criteria

Abstract : Multicriteria Decision Aiding (MCDA) is a field that aims at assisting expert decision mak ers (DM) in problems such as selecting, ranking, or classifying alternatives defined on several inter acting attributes. Such models do not make the decision, but assist the DM, who takes the final decision. It is thus crucial for the model to offer ways for the DM to maintain operational awareness, in particular in safety-critical contexts where errors can have dire consequences. It is thus a prerequisite of MCDA models to be intelligible, in terpretable, and to have a behaviour that is highly constrained by information stemming from in do main knowledge. Such models are usually built hand in hand with a field expert, obtaining infor mation through a Q&A procedure, and eliciting the model through methods rooted in operations research. On the other hand, Machine Learning (ML), and more precisely Preference Learning (PL), bases its approach on learning the optimal model from fitting data. This field usually focuses on model performances, tuning complex black-boxes to ob tain a statistically low error on new examples cases. While this is adapted to many settings, it is out of the question for decision aiding settings, as neither constrainedness nor intelligibility are available. This thesis bridges both fields. We focus on a certain class of MCDA models, called utilitaris tic hierarchical Choquet integrals (UHCI). Our first contribution, which is theoretical, is to show the identifiability (or unicity of the parameterization) of UHCIs This result motivates our second con tribution: the Neur-HCI framework, an archi tecture of neural network modules which can learn the parameters of a UHCI. In particular, all Neur HCI models are guaranteed to be formally valid, fitting the constraints that befit such a model, and remain interpretable. We show empirically that Neur-HCI models perform well on both artificial and real dataset, and that they exhibit remarkable stability, making it a relevant tool for alleviating the model elicitation effort when data is readily available, along with making it a suitable analysis tool for indentifying patterns in the data.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, March 4, 2022 - 9:20:08 AM
Last modification on : Friday, August 5, 2022 - 9:27:28 AM
Long-term archiving on: : Sunday, June 5, 2022 - 6:31:58 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03596964, version 1


Roman Bresson. Neural learning and validation of hierarchical multi-criteria decision aiding models with interacting criteria. Artificial Intelligence [cs.AI]. Université Paris-Saclay, 2022. English. ⟨NNT : 2022UPASG008⟩. ⟨tel-03596964⟩



Record views


Files downloads