System Effect Estimation by Sharding: A Comparison Between ANOVA Approaches to Detect Significant Differences

摘要

The ultimate goal of the evaluation is to understand when two IR systems are (significantly) different. To this end, many comparison procedures have been developed over time. However, to date, most reproducibility efforts focused just on reproducing systems and algorithms, almost fully neglecting to investigate the reproducibility of the methods we use to compare our systems. In this paper, we focus on methods based on ANalysis Of VAriance (ANOVA), which explicitly model the data in terms of different contributing effects, allowing us to obtain a more accurate estimate of significant differences. In this context, recent studies have shown how sharding the corpus can further improve the estimation of the system effect. We replicate and compare methods based on "traditional" ANOVA (tANOVA) to those based on a bootstrapped version of ANOVA (bANOVA) and those performing multiple comparisons relying on a more conservative Family-wise Error Rate (FWER) controlling approach to those relying on a more lenient False Discovery Rate (FDR) controlling approach. We found that bANOVA shows overall a good degree of reproducibility, with some limitations for what concerns the confidence intervals. Besides, compared to the tANOVA approaches, bANOVA presents greater statistical power, at the cost of lower stability. Overall, with this work, we aim at shifting the focus of reproducibility from systems alone to the methods we use to compare and analyze their performance.