NMF-based stochastic models for multi-label propagation

Multi-label propagation associates with transmitting multi-labels from tagged examples to untagged ones that are relevant in semantic. Many existing algorithms will lose their advantages if the correlations among tagged examples, untagged examples, and labels are improperly identified. To address this problem, a set of stochastic models are constructed, which describe not only the correlations among examples in the feature space, but also the correlations between examples and labels in the label space. The parameters involved in the stochastic models are then learned by a non-negative matrix factorization algorithm and a non-negative least square optimization algorithm. By setting up connections among examples, stochastic models, and labels, the advantages of smooth assumption are fully exploited, so that the labels are properly propagated. The experiments are carried out on six benchmark datasets from different real world applications. The results demonstrate the performance of the proposed approach, as compared with the performance of other state-of-the-art algorithms.