The complexity of resilience and responsibility for self-join-free conjunctive queries

Abstract

Several research thrusts in the area of data management have focused on understanding how changes in the data affect the output of a view or standing query. Example applications are explaining query results, propagating updates through views, and anonymizing datasets. These applications usually rely on understanding how interventions in a database impact the output of a query. An important aspect of this analysis is the problem of deleting a minimum number of tuples from the input tables to make a given Boolean query false. We refer to this problem as "the resilience of a query" and show its connections to the well-studied problems of deletion propagation and causal responsibility. In this paper, we study the complexity of resilience for self-join-free conjunctive queries, and also make several contributions to previous known results for the problems of deletion propagation with source side-effects and causal responsibility: (1) We define the notion of resilience and provide a complete dichotomy for the class of self-join-free conjunctive queries with arbitrary functional dependencies; this dichotomy also extends and generalizes previous tractability results on deletion propagation with source side-effects. (2) We formalize the connection between resilience and causal responsibility, and show that resilience has a larger class of tractable queries than responsibility.(3) We identify a mistake in a previous dichotomy for the problem of causal responsibility and offer a revised characterization based on new, simpler, and more intuitive notions. (4) Finally, we extend the dichotomy for causal responsibility in two ways: (a) we treat cases where the input tables contain functional dependencies, and (b) we compute responsibility for a set of tuples specified via wildcards.complexities of both deletion propagation with source side-effects and causal responsibility with minor modifications. Deletion propagation and existing results.Databases allow users to interact with data through views, which are often conjunctive queries. Views can be used to simplify complex queries, enforce access control policies, and preserve data independence for external applications. Of particular interest is how deletions in the input data affect the view (which is a trivial problem), but also how deletions in the view could be achieved by appropriately chosen deletions in the input data (which is far less trivial). Concretely, the problem of deletion propagation [Buneman et al. 2002;Dayal and Bernstein 1982] seeks a set Γ of tuples in the input tables that should be deleted from the database in order to delete a particular tuple from the view. Intuitively, this deletion should be achieved with minimal side-effects, where side-effects are defined with either of two objectives: (a) deletion propagation with source side-effects (DP source ) seeks a minimum set of input tuples Γ in order to delete a given output tuple; whereas (b) deletion propagation with view side-effects (DP view ) seeks a set of input tuples Γ that results in a minimum numb...

Propagation history ranking in social networks: A causality-based approach

“…Using this definition, Chockler and Halpern [19] introduced the degree of responsibility to evaluate the contribution of each cause. Recently, Meliou et al [21,49] first studied the complexity of causality and responsibility problems in relational databases. Qin et al [27,50] analyzed the causality problem for conjunctive queries with inequalities.…”

Themis: automatically testing software for discrimination

“…Causal relationships in data management systems [22,37,38] can help explain query results [40] and debug errors [53][54][55] by tracking and using data provenance [39]. For software systems that use data management, such provenance-based reasoning may aid testing for causal relationships between input attributes and outputs.…”

Software fairness

“…Causal relationships in data management systems [32,54,55] can help explain query results [59] and debug errors [79][80][81] by tracking and using data provenance [57]. For software systems that use data management, such provenance-based reasoning may aid testing for causal relationships between input attributes and outputs.…”