| dc.description.abstract | Combining multiple ranked lists of items, called voters, into a single consensus list is a popular problem with significant implications in numerous areas, including Bioinformatics, recommendation systems, metasearch engines, etc. Multiple recent solutions introduced supervised and unsupervised techniques that try to model the ordering of the list elements and identify common ranking patterns among the voters. Nevertheless, these works either require additional information (e.g. the element scores assigned by the voters, or training data), or they merge similar voters without the evidence that similar voters are important voters. Furthermore, these models are computationally expensive. To overcome these problems, this paper introduces an unsupervised method that identifies the expert voters, thus enhancing the aggregation performance. Specifically, we build upon the concept that collective knowledge is superior to the individual preferences. Therefore, the closer an individual list is to a consensus ranking, the stronger the respective voter is. By iteratively correcting these distances, we assign converging weights to each voter, leading to a final stable list. Moreover, to the best of our knowledge, this is the first work that employs these weights not only to assign scores to the individual elements, but also to determine their population. The proposed model has been extensively evaluated both with well-established TREC datasets and synthetic ones. The results demonstrate substantial precision improvements over three baseline and two recent state-of-the-art methods. © 2022 Elsevier Ltd | en |
