|
Abstract:
|
The rank aggregation problem is to combine ranking results from different sources in order to generate a "better" ranking list. Existing rank aggregation methods utilize only rank information. In some real applications, however, rank data come with covariates information. Moreover, some rankers are more informative and reliable than others in practice. Therefore, it is desirable to incorporate covariates information and distinguish high quality rankers from spam rankers. Here, we propose the Bayesian aggregation of rank-data with covariates (BARC) and its weighted version BARCW to overcome these limitations. Both BARC and BARCW employ latent variable models, which can incorporate covariates information, for ranked entities. Besides, BARCW quantitatively measures the difference in qualities of rankers. We generate aggregated ranking and its probability interval based on posterior distribution of latent variables estimated using Gibbs sampling with parameter expansion. Simulation studies show that our methods outperform the competing methods in different scenarios. Real-data applications validate our estimation of realiability parameters and effects of covaraiates.
|
