Robust Aggregation of Expert Opinions Based on Conflict Analysis and Resolution

This paper presents a new technique for combining the opinions given by a team of independent experts. Each opinion is associated with a confidence level that represents the expert’s conviction on its own judgement. The proposed technique first measures the conflict level introduced by every expert by taking into account the similarity between both its opinion and confidence, and those of the other experts within the team. An expert who disagrees with the majority of other experts with a similar confidence level is assumed to be conflicting (an “outlier” expert). Based on those conflict levels, an arbitration mechanism determines the reliability associated with each expert, by considering that a reliable expert is the one which is both confident and non-conflicting. Finally, the aggregated opinion is obtained as a weighted average (linear opinion pool) of the original expert opinions, with the weights being the reliability levels determined before. The proposed technique has been applied to texture image classification, leading to significantly better results than commonly-used opinion integration approaches.