Comparing metrics across TREC and NTCIR: the robustness to system bias

Test collections are growing larger, and relevance data constructed through pooling are suspected of becoming more and more incomplete and biased. Several studies have used evaluation metrics specifically designed to handle this problem, but most of them have only examined the metrics under incomplete but unbiased conditions, using random samples of the original relevance data. This paper examines nine metrics in a more realistic setting, by reducing the number of pooled systems. Even though previous work has shown that metrics based on a condensed list, obtained by removing all unjudged documents from the original ranked list, are effective for handling very incomplete but unbiased relevance data, we show that these results do not hold in the presence of system bias. In our experiments using TREC and NTCIR data, we first show that condensed-list metrics overestimate new systems while traditional metrics underestimate them, and that the overestimation tends to be larger than the underestimation. We then show that, when relevance data is heavily biased towards a single team or a few teams, the condensed-list versions of Average Precision (AP), Q-measure (Q) and normalised Discounted Cumulative Gain (nDCG), which we call AP', Q' and nDCG', are not necessarily superior to the original metrics in terms of discriminative power, i.e., the overall ability to detect pairwise statistical significance. Nevertheless, even under system bias, AP' and Q' are generally more discriminative than bpref and the condensed-list version of Rank-Biased Precision (RBP), which we call RBP'.