• Funding : Artois
  • Start year :
  • 2023

Description

The proposed subject is part of the current movement about hybrid AI, where efforts are made to develop AI systems based on both data and knowledge.

In this thesis, the main tasks of the systems to be considered are the classification task (predict the membership of an instance to a class, taken from a set of predefined classes) and the explanation of the classification carried out. The goal is to extract knowledge from data to complete the available pieces of knowledge and to exploit them to better classify / predict and also to better explain the classifications / predictions made.

Several knowledge-based approaches can be considered to get better predictions. For example, when pieces of knowledge describing logical links between sets of attributes are available, we can try to exploit them to clean the data before learning a predictor (by removing the instances that are in conflict or by correcting them when possible ), so as to improve the accuracy of the predictor. When the pieces of knowledge that are extracted consist of functional dependencies between attributes, dependent attributes can be removed before learning. When the pieces of knowledge that are extracted consist of classification rules deemed reliable and the predictor used is based on decision trees, this predictor can be corrected so that it complies with the rules. Depending on the approach used, the initial predictor will be transformed via new learning or by the implementation of a method for changing beliefs, without additional induction.

One of the objectives of this thesis is to compare such approaches and to determine in which situation it is interesting to use one approach rather than another, so as to improve the accuracy of the predictor obtained.

Another part of the thesis will focus on the exploitation of pieces of knowledge to better explain the classifications made. The knowledge extracted can, in fact, be used to simplify the abductive explanations of the classifications carried out (those which explain why the considered instance has been associated with a certain class by the predictor) but also to avoid considering impossible counter-examples when contrastive explanations are computed (those that explain why the considered instance has not been associated with a class other than its own by the predictor)…