<p>The dissemination of information at scale on social media has profound real-world consequences, necessitating advanced methods for cascade graph mining to identify and mitigate the spread of harmful content. Cascade classification, which involves categorizing information diffusion patterns based on their structural characteristics, is a fundamental task in this domain. While Contrastive Cascade Graph Learning (CCGL) has emerged as a promising self-supervised framework, its effectiveness and robustness for cascade classification remain under-explored, particularly in label-scarce scenarios. In this study, we provide a comprehensive evaluation of CCGL for classifying both synthetic cascades (generated by diverse network and diffusion models) and real-world cascades from multiple social media platforms. We further benchmark CCGL against representative baselines, including DeepHawkes and CasFlow. Our experimental results demonstrate that CCGL achieves strong performance, particularly on large-scale and structurally complex real-world cascades, where it outperforms representative baselines. However, the results also show that this advantage is conditional: sequence-based and cascade modeling baselines are competitive or superior in some settings, such as small citation-based cascades and synthetic network-model classification. Furthermore, we find that CCGL maintains high classification accuracy even with only 20% of labeled data in several real-world settings. We also show that synthetic pre-training can be beneficial, although its effectiveness depends on the target data distribution.</p>