In real-world scenarios, there are rich interaction relationships among users on different platforms such as e-commerce, consumer reviews, and social networks. Constructing these relationships into a graph structure and applying graph neural networks (GNNs) for malicious user detection has become a research trend in related fields in recent years. However, due to the small proportion of malicious users, as well as their disguising and high labeling cost, traditional GNN methods are limited by the problems of data imbalance, data inconsistency, and label scarcity. This study proposes a semi-supervised graph representation learning-based method for detecting malicious nodes. The method improves the GNN method for node representation learning and classification. Specifically, a class-aware malicious node detection (CAMD) method is constructed, which introduces class-aware attention mechanism, inconsistent GNN encoder, and class-aware imbalance loss functions to solve the problems of data inconsistency and imbalance. Furthermore, to address the limitation of CAMD in detecting malicious nodes with scarce labels, a graph contrastive learning-based method CAMD+ is proposed, which introduces data augmentation, self-supervised graph contrastive learning, and class-aware graph contrastive learning to enable the model to learn more information from unlabeled data and fully utilize scarce label information. Finally, a large number of experimental results on real-world datasets verify that the proposed methods outperform all baseline methods and show good detection performance in situations with different degrees of label scarcity.