To address the issues of poor separability and unclear intrinsic correlations of high-dimensional samples in existing data-driven transient stability assessment methods, a semi-supervised guided two-stage dimensionality reduction and visualization method is proposed. First, a pseudo-label generation method based on dual-index confidence and hierarchical clustering guidance is used to screen unlabeled samples with high confidence. Then, by incorporating a semi-supervised modification into the distance metric of modified t-distributed stochastic neighbor embedding (t-SNE), the method effectively separates samples of different categories while preserving the intrinsic structure of the data to the greatest extent. Finally, pairwise controlled manifold approximation projection (PaCMAP) is applied to map high-dimensional samples into a two-dimensional visualization space. The proposed method enables high-dimensional samples representing power system operating states and their stability to be projected into a 2D plane. For unknown operating conditions, their positions and stability status can be quickly determined based on the distribution of known operating point. Verification results on a large-scale system with tens of thousands of nodes and on a real-world power grid demonstrate the effectiveness of the proposed method.