A further step for efficient corrections of inconsistent probabilistic data sets

Abstract

Partial conditional probability assessments are having renewed attention and the merging of several sources of information is one of the more compelling needs associated with them. We focus here on the consequent task of correcting inconsistent probabilistic databases.We propose an efficient method for correcting incoherent (i.e. inconsistent) conditional probability assessments, that has a polynomial space complexity, differently from methods based on probabilistic satisfiability problems (PSAT) which require an exponential amount of memory space.This method uses Mixed Integer Programming (MIP) procedure to minimize the L1 distance between probability assessments and exploits the presence of the so-called “zero layers”. Through a simple prototypical example, we illustrate the feasibility and the peculiarities of the proposed procedure.Finally, we show some experimental results obtained through randomly generated incoherent assessments.