We conduct extensive experiments on two new real-world KG datasets, i.e., Freebase A large portion of books are labeled into eight genres of literature. Each labeled book has only one label. Compared with node or edge classification, aggregating node embeddings for graph-level classification needs more context information. Therefore, it is difficult to find a universal representation learning approach that solves all different levels of graph mining tasks. We've conducted a preliminary analysis on aggregating our learned node embeddings for graph-level The results on the popular MUT AG dataset are listed in Table 2. LM-based models (e.g., LM (XRoBERTa), LM (GPT -2), and LM (DRoBERTa)) are able to learn accurate and rich node attributes, leading to superior performance in node classification.