A Joint Maximum Likelihood Estimation Framework for Truth Discovery: A Unified Perspective

Truth discovery algorithms have been widely applied to identify the true claims from the conflicting information provided by multiple sources. In general, they conduct an iterative procedure to estimate source reliability degrees as weights and infer the true claims via weighted voting. However, there is little prior work that provides theoretical analysis on the convergence of truth discovery methods. In this paper, we formulated the truth discovery task as a joint maximum likelihood estimation (JMLE) problem for unknown source reliability and truth claims. Within this framework, we proposed a Unified Truth Discovery (UTD) algorithm to get the numerical solution to JMLE for truth and source reliability. With mild conditions, we proved the consistency of the JMLE and the convergence of the proposed UTD algorithm. In addition, our proposed UTD algorithm turns out to include many existing truth discovery algorithms as special cases. This guarantees that our theoretical results can be applied to these truth discovery algorithms. We further conduct extensive experiments on synthetic data sets as well as five real-world data sets, and results from these numerical analysis support the theoretical results of the proposed UTD algorithm and the other state-of-the-art truth discovery algorithms.

Automated summary

This paper proposes a unified truth discovery algorithm, which uses maximum likelihood estimation to estimate source reliability and truth values. It proves the consistency of the estimation and the convergence of the algorithm, and conducts experiments to support the theoretical results.