Causality in Data Management: Foundations and Applications

点击翻译按钮获取中文摘要

Trying to understand why and how the occurrence of an event or the execution of an action affect other events or properties of objects belongs to the essence of human existence. If the search for causal explanations is not the oldest activity performed by humans, it is certainly one of them, yet it still has been an elusive concept that has been studied in many fields of knowledge, ranging from philosophy to computer science.******In these times of big data and data science, analyzing large volumes of data has become an increasingly common and complex problem. Beyond the traditional problem of extracting explicit data from a data source, it has become crucial to understand and make sense of the data, and extract implicit knowledge from such a source. In particular, finding explanations for data phenomena as shown, for example, in query answers, satisfaction or violation of semantic constraints, and view contents, has been identified as an important task, for which explanation models, algorithmic results, and practical computational implementations still need to be developed. Causality is used to address the problem of providing explanations for different forms of manifestations of data.******Although causality has been investigated in several areas, e.g. statistics, artificial intelligence, economics, its emergence in data management, and databases in particular, is quite recent. Addressing causality in databases requires a mathematical characterization of the notion of cause, the investigation of the mathematical model, and the algorithmic and complexity analysis of computational problems. Among the latter, we find development of efficient algorithms for computing causes for query answers, whenever possible, or of efficient approximation algorithms when the intrinsic complexity of the problem is high. It also becomes necessary to rank causes, identifying those most relevant, for which the notion of responsibility has been introduced. Several computational problems emerge around responsibility computation. The PI has already obtained interesting results in this area. However, many aspects and problems of causality in databases, and more generally in data management, are still open.******In this proposal we address causality in the context of semantically enriched data sources, with the aim to propose and investigate a model of causality in scenarios such as ontology-based data access, virtual data integration, and multidimensional databases, among others. We also go into more fundamental problems, such as the logical characterization of causes, which should tell us what theory embedding the data source can be used to infer causes and how. We investigate the problem of learning causal relations from (possibly probabilistic) data sources, by developing machine learning methods specifically tailored for databases, which normally lack the additional semantic information that is useful for learning.