SourceVote: Fusing multi-valued data via inter-source agreements

Abstract

Data fusion is a fundamental research problem of identifying true values of data items of interest from conflicting multi-sourced data. Although considerable research efforts have been conducted on this topic, existing approaches generally assume every data item has exactly one true value, which fails to reflect the real world where data items with multiple true values widely exist. In this paper, we propose a novel approach, SourceVote, to estimate value veracity for multi-valued data items. SourceVote models the endorsement relations among sources by quantifying their two-sided inter-source agreements. In particular, two graphs are constructed to model inter-source relations. Then two aspects of source reliability are derived from these graphs and are used for estimating value veracity and initializing existing data fusion methods. Empirical studies on two large real-world datasets demonstrate the effectiveness of our approach.