<p>As a typical semi-supervised classification method, label propagation has gained widespread attention in recent years. Existing label propagation methods generally consist of two fundamental steps: constructing an affinity graph and formulating propagation rules. While most methods focus on designing sophisticated propagation rules derived from physical theorems or statistical principles, little attention has been paid to the critical aspect of affinity graph learning which may affect the classification performance. This study aims to improve the semi-supervised classification performance of label propagation from a new perspective–multi-view affinity graph learning. Due to uncertainties in the initial sample distribution, constructing the affinity graph from a single view may fail to fully capture the propagation probability between samples. To address this, we propose an enhanced label propagation method named ELP-TAF, which evaluates the affinity between sample pairs from three different views and generates the final propagation matrix through a multi-view graph fusion technique. Extensive experiments on both synthetic and real benchmark datasets demonstrate the competitive performance of ELP-TAF in semi-supervised classification tasks. Moreover, the application of high-performance computing and parallel processing may unlock the potential of ELP-TAF for addressing large-scale real-world problems.</p>