Estimating probability values from an incomplete dataset

AbstractAn essential component in Machine Learning processes is to estimate any uncertainty measure reflecting the strength of the relationships between variables in a dataset. In this paper we focus on those particular situations where the dataset has incomplete entries, as most real-life datasets have. We present a new approach to tackle this problem. The basic idea is to initially estimate a set of probability intervals that will be used to complete the missing values. Then, these values are used to obtain new bounds of the expected number of entries in the dataset. The probability intervals are narrowed iteratively until convergence. We have shown that the same processes can be used to estimate both, probability intervals and probability distributions, and give conditions that guarantee that the estimator is the correct one