TADG: topology-aware and distillation-guided framework for continual knowledge graph embedding

When dealing with dynamically evolving knowledge graphs, traditional knowledge graph embedding methods lack an efficient incremental update mechanism. They typically require retraining on the entire graph and cannot effectively reuse previously learned knowledge. Existing continual knowledge graph embedding (CKGE) approaches attempt to address this issue, but they still have notable limitations. First, strategies that freeze historical knowledge and fine-tune only newly added facts can reduce training cost, yet they sever the connection between old and new knowledge and fail to capture the semantic evolution of existing knowledge under new contexts. Second, regularization-based methods aim to stabilize model parameters while integrating new knowledge, but their updates are usually restricted to entities and relations directly connected to the newly introduced knowledge, lacking the ability to model coordinated evolution over the global graph structure. To overcome these limitations, this paper proposes a continual learning framework driven by topology-aware and distillation guidance. The goal is to efficiently learn new knowledge while identifying low-interference update directions that encourage coordinated evolution of historical representations, thereby balancing the acquisition of new knowledge with the preservation of prior knowledge. The framework first performs fast local updates to learn newly added knowledge. It then adjusts the strength of constraints on old knowledge by quantifying adjacency information entropy, enabling finer-grained protection of critical knowledge and effectively mitigating catastrophic forgetting. Finally, it employs distillation-guided gradient projection to steer updates toward directions that reduce deviation from the previous model on representative historical memories, enabling more stable and coordinated knowledge updating. Extensive experiments on multiple public datasets demonstrate that the proposed method outperforms continual learning baselines under a unified backbone and evaluation protocol.