TAMPI: A Time-aware Multi-perspective Interaction Framework for Temporal Knowledge Graph Completion

The interactions of digital virtual humans in social media circumstance, can be viewed as a Knowledge Graph (KG). As the structure of KGs may vary over time, static Knowledge Graph Completion (KGC) methods do not deal with time-varying KGs. Recent research on temporal KGC (TKGC) task shows that the inclusion of time information can improve the performances of KG’s embedding models on temporal KGs. However, the current TKGC models commonly internalizes time features into the embeddings of entities and relations, which reduces the dimensionality of quadruples to triples and only generate the representation vectors for entities and relations, and the information embedded in the time perspective is regretfully not represented and modeling. Towards this challenge, we propose a time-aware multi-perspective interaction (TAMPI) framework for TKGC task, which follows an encoder-decoder architecture. Especially, this work innovatively integrates time features into other perspectives in addition to time-independently modeling, to fully represent all the dimensions of the given quadruples and achieve better TKGC efficiency. The proposed framework can provide more accurate temporal perspective features, and improve the performance of completing temporal KGs. Experiments show the benefits of our approach on temporal KGs.