Random Walk in Feature-Sample Networks for Semi-supervised Classification

Positive-unlabeled learning is a semi-supervised task in which only some positive-labeled and many unlabeled samples are available. The goal of its transductive setting is to label all unlabeled data at once. In this paper, we developed a technique to grade positive-class pertinence levels of each sample, and we interpret the grades to classify the unlabeled ones. In our method, a sparse binary matrix represents the input data, which determines the feature-sample network whose vertices represent samples and attributes. The limiting probabilities of a random walk in the network estimate the pertinence levels. The results are evaluated regarding both class discrimination and classification accuracy. Computer simulations reveal that our model performs well in positive-unlabeled learning, especially with few labeled samples. Notably, the outcomes compare to the results from supervised methods, which profit from most data labeled. Additionally, the technique has linear time and space complexity if the input dataset is already in a sparse representation. The low computational cost of the construction and update of the feature-sample network allows for extensions of the technique to several learning problems, including online learning and dimensionality reduction.