As the feature size keeps shrinking in the modern semiconductor manufacturing process, subresolution assist feature (SRAF) insertion is one promising resolution enhancement technique that can improve the printability and lithographic process window of target patterns. Model-based SRAF generation achieves a high accuracy but with a high computational cost, while rule-based SRAF insertion may require a huge rule table to handle complex patterns. Thus, state-of-the-art works resort to machine learning to reduce runtime but require abundant training samples to generalize the trained models and achieve high performance. Nevertheless, in advanced lithography, we may have a huge solution space of SRAF insertion but few labeled training samples. Therefore, in this work, we address SRAF insertion from a data efficiency perspective. We separate sample selection from SRAF probability learning and train a variational autoencoder and an adversarial network to discriminate between unlabeled and labeled data effectively. Second, we devise a region-based concentric circle area sampling representation to avoid information loss during feature extraction. Third, we determine the final placement of SRAFs by a novel clustering method based on retrieved data points. Experimental results show that, compared with state-of-the-art works, by using 40% training samples, our framework can achieve comparable or even better process variation bands and edge placement errors.