Comparing top-k XML lists

Systems that produce ranked lists of results are abundant. For instance, Web search engines return ranked lists of Web pages. There has been work on distance measure for list permutations, like Kendall tau and Spearman's footrule, as well as extensions to handle top-k lists, which are more common in practice. In addition to ranking whole objects (e.g., Web pages), there is an increasing number of systems that provide keyword search on XML or other semistructured data, and produce ranked lists of XML sub-trees. Unfortunately, previous distance measures are not suitable for ranked lists of sub-trees since they do not account for the possible overlap between the returned sub-trees. That is, two sub-trees differing by a single node would be considered separate objects. In this paper, we present the first distance measures for ranked lists of sub-trees, and show under what conditions these measures are metrics. Furthermore, we present algorithms to efficiently compute these distance measures. Finally, we evaluate and compare the proposed measures on real data using three popular XML keyword proximity search systems.