Online Semi-supervised Pairwise Learning

Online learning is a machine learning method that sequentially updates the predictive model. It is a significant learning technique for massive and streaming data, where it is impractical to store the data for training. For such large-scale real-world data, it is also infeasible to label the whole training samples. Online semi-supervised learning concerns learning a model on both labeled and unlabeled examples of streaming data. The online semi-supervised pairwise learning optimizes an objective function in which its loss function is based on pairs of examples. The recent online semi-supervised pairwise learning method builds a first-order pairwise classifier that lacks the generalization ability of batch semi-supervised methods. To improve the generalization capacity of the pairwise model, we propose a second-order online semi-supervised pairwise learning algorithm that exploits second-order information of the features. More specifically, we adopt a confidence-weighted model and reformulate its objective function for pairwise semi-supervised learning. The model treats unlabeled data as positive and negative while pairing the current example with the previous opposite one in building the model. Experiments on different benchmark and real-world datasets show that the proposed model achieves AUC results that surpass the existing state-of-the-art online semi-supervised method. Also, the proposed method shows a comparable AUC results to the batch semi-supervised method.