Recently, several large Knowledge Bases (KBs) have been constructed by mining the Web for information. As an increasing amount of inconsistent and non-reliable data are available, KBs facts may be uncertain and are then associated with an explicit certainty degree. When querying these uncertain KBs, users seek high quality results i.e., results that have a certainty degree greater than a given threshold α. However, as they usually have only a partial knowledge of the KBs contents, their queries may be failing i.e., they return no result for the desired certainty. To prevent this frustrating situation, instead of returning an empty set of answers, our approach explains the reasons of the failure with a set of αMinimal Failing Subqueries (αMFSs), and computes alternative relaxed queries, called αMaXimal Succeeding Subqueries (αXSSs), that are as close as possible to the initial failing query. Moreover, as the user may not always be able to provide an appropriate threshold α, we propose two algorithms to compute the αMFSs and αXSSs for other thresholds. Our experiments on the WatDiv benchmark show the relevance of our algorithms compared to a baseline method.

Several large uncertain Knowledge Bases (KBs) are available on the Web where facts are associated with a certainty degree. When querying these uncertain KBs, users seek high quality results i.e., results that have a certainty degree greater than a given threshold α. However, as they usually have only a partial knowledge of the KB contents, their queries may be failing i.e., they return no result for the desired certainty level. To prevent this frustrating situation, instead of returning an empty set of answers, our approach explains the reasons of the failure with a set of αMinimal Failing Subqueries (αMFSs), and computes alternative relaxed queries, called αMaXimal Succeeding Subqueries (αXSSs), that are as close as possible to the initial failing query. Moreover, as the user may not always be able to provide an appropriate threshold α, we propose three algorithms to compute the αMFSs and αXSSs for other thresholds, which also constitutes a relevant feedback for the user. Multiple experiments with the WatDiv benchmark show the relevance of our algorithms compared to a baseline method.