Graph Based Classification Methods Using Inaccurate External Classifier Information

摘要

In this paper we consider the problem of collectively classifying entitieswhere relational information is available across the entities. In practiceinaccurate class distribution for each entity is often available from another(external) classifier. For example this distribution could come from aclassifier built using content features or a simple dictionary. Given therelational and inaccurate external classifier information, we consider twograph based settings in which the problem of collective classification can besolved. In the first setting the class distribution is used to fix labels to asubset of nodes and the labels for the remaining nodes are obtained like in atransductive setting. In the other setting the class distributions of all nodesare used to define the fitting function part of a graph regularized objectivefunction. We define a generalized objective function that handles both thesettings. Methods like harmonic Gaussian field and local-global consistency(LGC) reported in the literature can be seen as special cases. We extend theLGC and weighted vote relational neighbor classification (WvRN) methods tosupport usage of external classifier information. We also propose an efficientleast squares regularization (LSR) based method and relate it to informationregularization methods. All the methods are evaluated on several benchmark andreal world datasets. Considering together speed, robustness and accuracy,experimental results indicate that the LSR and WvRN-extension methods performbetter than other methods.