Positive-Unlabeled Learning for Network Link Prediction

Link prediction is an important problem in network data mining, which is dedicated to predicting the potential relationship between nodes in the network. Normally, network link prediction based on supervised classification will be trained on a dataset consisting of a set of positive samples and a set of negative samples. However, well-labeled training datasets with positive and negative annotations are always inadequate in real-world scenarios, and the datasets contain a large number of unlabeled samples that may hinder the performance of the model. To address this problem, we propose a positive-unlabeled learning framework with network representation for network link prediction only using positive samples and unlabeled samples. We first learn representation vectors of nodes using a network representation method. Next, we concatenate representation vectors of node pairs and then feed them into different classifiers to predict whether the link exists or not. To alleviate data imbalance and enhance the prediction precision, we adopt three types of positive-unlabeled (PU) learning strategies to improve the prediction performance using traditional classifier estimation, bagging strategy and reliable negative sampling. We conduct experiments on three datasets to compare different PU learning methods and discuss their influence on the prediction results. The experimental results demonstrate that PU learning has a positive impact on predictive performances and the promotion effects vary with different network structures.