Fast semi-supervised classification based on anchor graph

Semi-supervised learning (SSL) is a popular field due to its ability to leverage fusion information of an extensive volume of unlabeled data alongside a modest amount of labeled data, which can fully use the given information. However, with the expanding data, traditional graph-based semi-supervised learning (GSSL) is unsuitable for handling large-scale problems owing to its high computational complexity. To solve this problem, a fast semi-supervised classification method based on anchor graph (FSSCAG) is proposed. FSSCAG combines a similarity graph with the learning of anchor affiliation matrix to get the data indicator matrix and the objective function can lead to clear classification results. First, we employ a neighbor assignment approach without parameters to build a similarity graph between data and anchors, and then solve the affiliation matrix of anchors by constructing a quadratic programming model. After that, a projected gradient method is used to solve our model. In this way, the learning of the data indicator matrix is transformed into the learning of the anchor affiliation matrix, which reduces the time complexity greatly. Results from experiments on real-world datasets demonstrate that FSSCAG reduces time expenditure while achieving comparable or even superior classification performance when compared to advanced algorithms.