By Paolo Franchi
On reasoning with partially reliable and conflicting sources. Ben Abdallah, Nadia; Jousselme, Anne-Laure. CMRE-FR-2019-004. September 2020.
When combining information provided by several sources such as intelligence reports, experts, sensors, or algorithms, inconsistencies between pieces of information provided by the sources are likely to arise reflecting the partial quality of the sources. A key step in developing intelligent automated systems that properly manage conflict is the consideration in the reasoning of knowledge on source quality. This report addresses questions related to the problem of reasoning with partially reliable and conflicting sources by shedding light on the relationships between the source characteristics and the message content. The main models and concepts are exemplified on a problem of vessel destination prediction. We investigate models of partially reliable sources in the probabilistic and evidential frameworks. We specifically detail models in the evidence theory setting, which offers a great flexibility in capturing dimensions of information quality such as the conflict, uncertainty and imprecision. We focus on the particular credibility quality dimension that we define with respect to the conflict between pieces of information. We provide a set of measures of credibility and describe a general family of credibility-based discounting models which extend some existing models by capturing a more general definition of the conflict. We finally investigate the theoretical foundation of the credibility-based models. More specifically, we address the question of reliability conduciveness of the credibility, i.e., if the relative lack of agreement between several testimonies brings information on the reliability of the sources. We first prove the existence of a particular credibility-based discounting model that is totally information-driven. Contrary to other models, this particular credibility-based discounting model does not make any additional assumption. However, it appears to be less informative on the reliability as it provides evidence solely on the likelihood of the report to be unreliable, and none on its likelihood to be reliable. This suggests that the conflict is not reliability-conductive, but rather unreliability-conductive.