Comparison of uncertainty representations for missing data in information retrieval

The aim of this paper is to assess the impact of the uncertainty representation for retrieving items with missing data. The problem of information retrieval from incomplete incident databases is addressed in this paper. After a brief survey on the problem of missing data with an emphasis on the information retrieval application, we propose a novel approach for retrieving case records with missing data. The general idea of the proposed data driven approach is to model the uncertainty pertaining to this missing data. We chose the general model of belief functions as it encompasses as special cases both classical and probability models. Several uncertainty models are then compared based on (1) an expressiveness criterion (non-specificity or randomness) and (2) objective measures of performance typical to the Information Retrieval domain. The results are illustrated on several datasets and a simulation controlled missing data mechanism.