An AI approach to relational data models for uncertain and imprecise information

Real world information is often imprecise and uncertain. While uncertainty management has been a central research issue in the AI community, the database community has been slow to adapt to this trend. There have been attempts to use uncertainty calculi developed in AI to data modeling for incomplete information: data models based on fuzzy set theory are considered good for data which are intrinsically lexically imprecise, but generally it is difficult to justify the usage of fuzzy membership functions for uncertain information. Data models based on Bayesian probability theory have well defined theoretical foundations, but the assumption of the availability of probability distributions for uncertain data is unreasonable in many practical applications. The Dempster-Shafer theory as a generalization of the Bayesian theory is an attempt to overcome those weaknesses.; This is the first approach to use the Dempster-Shafer theory in database applications. The Dempster-Shafer theory is applied to relational data modeling for uncertain and imprecise information. Since a basic probability assignment in the Dempster-Shafer theory, natural to represent ignorance as well as uncertainty, is used for the representation of an uncertain attribute value in a relation. The Bel and Pls functions in the Dempster-Shafer theory are extended at the tuple level to define Selection operation as well as Join, Union, Intersect and Projection. In this model, every tuple in a relation has a special attribute CL which represents our belief of the support for the tuple.; We identify two new problems in our model: the potential existence of identical tuples with different degrees of belief and the potential existence of conflicting data from different sources. The former is solved by the proposed relational algebra. One proposed solution to the latter is based on Dempster's rule of combination in the Dempster-Shafer theory, which performs pooling data from different sources for purpose of making choices between hypotheses. Dempster's rule is extended to define the evidential union and projection operations, for purpose of pooling data from different sources at the tuple level, and as a means of removing inconsistency in relations.