Autonomous driving systems frequently encounter the issue of missing data, which complicates accurate prediction and decision-making processes. To handle this problem, this paper investigates the uses of the Self-Training and Active Learning (STAL) method, a novel imputation approach designed specifically for graph-based machine learning. STAL enhances Graph Neural Networks (GNNs) by combining active learning (AL) and self-training (ST), thereby improving label efficiency and overall performance. By selectively labeling highly uncertain nodes and generating reliable pseudo-labels, STAL maximizes the use of unlabeled data during model training. Experimental results demonstrate significant improvements in node classification tasks, highlighting the effectiveness of STAL in reducing labeling costs while maintaining high accuracy.