Details

Title:Evaluating MLNs for Collective Classification
Authors:Crane, Robert J.
Serial Number:2010-04
Publication Date:12-13-2010
Abstract:Collective Classification is the process of labeling instances in a graph using both instance attribute information and information about relations between instances. While several Collective Classification Algorithms have been well studied, the use of Markov Logic Networks (MLNs) remains largely untested. MLNs pair first order logic statements with a numerical weight. With properly assigned weights, these rules may be used to infer class labels from evidence stated as logic statements. Our study evaluated MLNs against other Collective Classification algorithms on both synthetic data and real data from the CiteSeer dataset. As a whole, we encountered inconsistent and often poor performance with MLNs, especially on synthetic data where other Collective Classification algorithms performed well.
View ReportView bibtex