AdaTKG: Adaptive Memory Enhances Temporal Knowledge Graph Reasoning
AdaTKG presents an innovative strategy for reasoning with temporal knowledge graphs (TKGs) by conceptualizing each entity as an adaptive process that updates its memory with every interaction. In contrast to traditional approaches that rely on fixed entity representations, AdaTKG dynamically enhances these representations through a learnable exponential moving average controlled by a single shared scalar, thus removing the necessity for individual learnable parameters for each entity. This mechanism allows for online memory accumulation, leading to improved predictions as interactions increase. The research, available on arXiv (2605.07121), marks a shift from static representations, facilitating more effective management of evolving events within TKGs.
Key facts
- AdaTKG models entities as adaptive processes with per-entity memory.
- Memory updates occur with every observed interaction.
- Uses a learnable exponential moving average with a single shared scalar.
- Eliminates per-entity learnable parameters.
- Improves predictions as more interactions arrive.
- Published on arXiv with ID 2605.07121.
- Addresses limitations of static entity representations in TKGs.
- Focuses on temporal knowledge graph reasoning over evolving events.
Entities
Institutions
- arXiv