This paper presents a semi-supervised learning framework based on graph embeddings. The framework jointly trains embeddings for each instance to predict both the class label and the neighborhood context in the graph. Two variants of the method are developed: a transductive variant, which uses the graph structure to predict labels of instances already present in the graph, and an inductive variant, which defines embeddings as a parametric function of the feature vectors to enable predictions on unseen instances. The method is evaluated on benchmark tasks including text classification, distantly supervised entity extraction, and entity classification, where it outperforms existing models. The transductive variant achieves up to 8.5% improvement over other methods, while the inductive variant achieves up to 18.7% improvement. The framework combines supervised and unsupervised learning objectives, using graph embeddings to predict both class labels and graph context. The method is shown to be effective in both transductive and inductive settings, with the inductive variant leveraging feature vectors to improve performance. The results demonstrate that joint training of classification and graph context prediction leads to significant improvements over existing methods.This paper presents a semi-supervised learning framework based on graph embeddings. The framework jointly trains embeddings for each instance to predict both the class label and the neighborhood context in the graph. Two variants of the method are developed: a transductive variant, which uses the graph structure to predict labels of instances already present in the graph, and an inductive variant, which defines embeddings as a parametric function of the feature vectors to enable predictions on unseen instances. The method is evaluated on benchmark tasks including text classification, distantly supervised entity extraction, and entity classification, where it outperforms existing models. The transductive variant achieves up to 8.5% improvement over other methods, while the inductive variant achieves up to 18.7% improvement. The framework combines supervised and unsupervised learning objectives, using graph embeddings to predict both class labels and graph context. The method is shown to be effective in both transductive and inductive settings, with the inductive variant leveraging feature vectors to improve performance. The results demonstrate that joint training of classification and graph context prediction leads to significant improvements over existing methods.