General-Domain Truth Discovery via Average Proximity

Truth discovery is a general name for a broad range of statistical methods aimed to extract the correct answers to questions, based on multiple answers coming from noisy sources. For example, workers in a crowdsourcing platform. In this paper, we suggest a simple heuristic for estimating workers' competence using average proximity to other workers. We prove that this estimates well the actual competence level and enables separating high and low quality workers in a wide spectrum of domains and statistical models. We then design a simple proximity-based truth discovery algorithm (\PTD) that weighs workers according to their average proximity. The answers for questions may be of different forms such as real-valued, categorical, rankings, or other complex labels, and \PTD can be combined with any existing aggregation function or voting rule to improve their accuracy. We demonstrate through an extensive empirical study on real and synthetic data that \PTD and its iterative variants outperform other heuristics and state-of-the-art truth discovery methods in the above domains.