Building Test Collections using Bandit Techniques: A Reproducibility Study

Abstract

The high cost of constructing test collections led many researchers to develop intelligent document selection methods to find relevant documents with fewer judgments than the standard pooling method requires. In this paper, we conduct a comprehensive set of experiments to evaluate six bandit-based document selection methods, in terms of evaluation reliability, fairness, and reusability of the resultant test collections. In our experiments, the best performing method varies across test collections, showing the importance of using diverse test collections for an accurate performance analysis. Our experiments with six test collections also show that Move-To-Front is the most robust method among the ones we investigate. © 2020 ACM.