An Iterative Learning Algorithm for Within-Network Regression in the Transductive Setting

Within-network regression addresses the task of regression in partially labeled networked data where labels are sparse and continuous. Data for inference consist of entities associated with nodes for which labels are known and interlinked with nodes for which labels must be estimated. The premise of this work is that many networked datasets are characterized by a form of autocorrelation where values of the response variable in a node depend on values of the predictor variables of interlinked nodes. This autocorrelation is a violation of the independence assumption of observation. To overcome to this problem, the lagged predictor variables are added to the regression model. We investigate a computational solution for this problem in the transductive setting, which asks for predicting the response values only for unlabeled nodes of the network. The neighborhood relation is computed on the basis of the node links. We propose a regression inference procedure that is based on a co-training approach according to separate model trees are learned from both attribute values of labeled nodes and attribute values aggregated in the neighborhood of labeled nodes, respectively. Each model tree is used to label the unlabeled nodes for the other during an iterative learning process. The set of labeled data is changed by including labels which are estimated as confident. The confidence estimate is based on the influence of the predicted labels on known labels of interlinked nodes. Experiments with sparsely labeled networked data show that the proposed method improves traditional model tree induction.