The presentation describes a semi-supervised framework based on a fast and efficient feature representation with a highly scalable learning approach, achieving high accuracy and substantial gains in computational cost. The framework makes possible to implement a semi-supervised approach in large-scale settings. More specifically, the framework combines the VLAD (Vector of Locally Aggregated Descriptors) feature aggregation method and PCA (Principal Component Analysis) for image representation, and proposes the use of Approximate Laplacian Eigenmaps (ALEs) for learning concepts in time linear to the number of images. Extending the framework, we construct an incremental method that is applicable in online problems, where the learning model has to be updated once a new image is observed. With this extension, we can process data from social networks and the Web in real-time, exploiting the impact of unknown data.