Contrastive Learning with Growing Generated Representations for Inductive Knowledge Graph Embedding

Inductive Knowledge Graphs (KGs) operate in a constrained, dynamically evolving setting where the target KG is derived from the source KG under a fixed relational schema. Representation learning on inductive KGs is considerably more challenging than in the static transductive setting because the model must extract relational patterns from the source KG and robustly generalize them to unseen entities in the target KG. However, existing methods remain constrained by 1) sparsity which hampers robust pattern extraction, 2) implicit transfer which limits effective generalization across source and target KGs, and 3) the lack of systematic scaling analysis, leaving unclear how model capacity shapes inductive transfer. In this paper, we first propose VMCL, a Contrastive Learning (CL) framework with graph-guided Variational Autoencoder on Meta-KGs, which effectively alleviates the limitations caused by sparsity and implicit transfer in the inductive setting. Then, we conduct a systematic scaling study of VMCL by increasing the dimension of representations, the number of meta-KGs, and the number of generated representations. Extensive experiments demonstrate that VMCL consistently outperforms state-of-the-art baselines and exhibits scaling benefits as model capacity increases.