How to evaluate rankings of academic entities using test data

Abstract In the field of scientometrics, impact indicators and ranking algorithms are frequently evaluated using unlabelled test data comprising relevant entities (e.g., papers, authors, or institutions) that are considered important. The rationale is that the higher some algorithm ranks these entities, the better its performance. To compute a performance score for an algorithm, an evaluation measure is required to translate the rank distribution of the relevant entities into a single-value performance score. Until recently, it was simply assumed that taking the average rank (of the relevant entities) is an appropriate evaluation measure when comparing ranking algorithms or fine-tuning algorithm parameters. With this paper we propose a framework for evaluating the evaluation measures themselves. Using this framework the following questions can now be answered: (1) which evaluation measure should be chosen for an experiment, and (2) given an evaluation measure and corresponding performance scores for the algorithms under investigation, how significant are the observed performance differences? Using two publication databases and four test data sets we demonstrate the functionality of the framework and analyse the stability and discriminative power of the most common information retrieval evaluation measures. We find that there is no clear winner and that the performance of the evaluation measures is highly dependent on the underlying data. Our results show that the average rank is indeed an adequate and stable measure. However, we also show that relatively large performance differences are required to confidently determine if one ranking algorithm is significantly superior to another. Lastly, we list alternative measures that also yield stable results and highlight measures that should not be used in this context.