Graph-induced rank-aggregation using information fusion operators

Rank-aggregation, the process of fusing multiple ranked lists into a single unified ranking, is a cornerstone of various information retrieval systems. Given the widespread adoption of these techniques and the increasing demand for precise and relevant results, the development of innovative rank aggregation algorithms is imperative. This paper proposes a novel rank-aggregation algorithm which leverages graph theory and information fusion to model inter-element relationships and combine ranked lists effectively. By constructing a rankers’ similarity graph and employing spectral clustering, the algorithm identifies top-performing rankers, whose rankings are combined using the random forest algorithm, which is applied several times to produce different combined rankings and score them based on the average expertise of included rankers. Finally, information fusion theory is employed to aggregate the ensemble of rankers’ outputs through the optimistic and pessimistic exponential ordered weighted averaging operators. Experimental results on LETOR4.0 datasets demonstrate noticeable improvements compared to baselines, highlighting the efficiency of the proposed algorithm using only a limited number of rankers.