Improved Unbiased Node Label Prediction Algorithm
Yu Gan · 2015
In social networks,predictions of attributes and locations of users and labels of images are extensively applied in many fields.In order to improve the performance of label prediction,this paper proposed an improved unbiased node label prediction algorithm.Firstly,we formalized the label prediction problem in social networks.Secondly,based on the mismatch of the maximization of joint likelihood of training objective under all observed labels and the single variable marginal prediction scores conditioned by the observed labels,we proposed an improved graphical model training algorithm.Finally,according to the unbiased estimation of confidence,we proposed a training model not including additional labels based on sub-graph method.Experiments on the Twitter and Pokec datasets show that,compared with related works,the proposed algorithm has better accuracy and execution efficiency while predicting labels.