Dual-layer Semantic Propagation for Semi-supervised Projection

Graph-based semi-supervised projection methods improve the dimension reduction performance by performing the label propagation among the labeled and unlabeled samples. However, most of the existing methods only consider the single-layer semantic propagation on the graph and adopt a mandatory semantic consistent constraint that forces the learned pseudo-labels of labeled samples to be directly equal to the given labels, which may bring semantic loss and thus reduce the subsequent projection performance. To solve the problems, we propose a Dual-layer Semantic Propagation for Semi-supervised Projection (DSP-SP). Our model can simultaneously perform indirect semantic propagation on a structured graph, direct semantic propagation with a soft semantic alignment and semi-supervised projection in a unified semi-supervised learning framework. Specifically, the soft semantic alignment reduces the semantic loss by making the pseudo-labels approximate the given labels as possible, which is beneficial to learning feature projection matrix. Experiments show the superiority of the proposed method on both projection accuracy and learning efficiency.