Adaptive Decay Model Improves Knowledge Graph Retrieval
A recent study published on arXiv introduces a hierarchical model for knowledge graphs, which substitutes uniform decay with a continuous decay surface defined by velocity and volatility. The researchers contend that various types of knowledge display distinct temporal dynamics, with the primary challenge during retrieval being the identification of significant information at the time of the query. This framework comprises three levels of learnable parameters: domain-level, context-level, and entity-level.
Key facts
- Knowledge graphs used for retrieval treat all facts as equally current.
- Existing temporal approaches apply uniform decay using a single forgetting curve.
- The paper argues this is fundamentally misspecified.
- Different knowledge types exhibit different temporal dynamics.
- The core retrieval problem is identifying what is important at query time.
- The proposed framework replaces uniform decay with a continuous decay surface.
- The decay surface is parameterized by velocity and volatility.
- Velocity measures how frequently a concept is observed.
- Volatility measures how much the value changes between observations via embedding distance.
- The decay surface is decomposed into three learnable levels: domain-level, context-level, and entity-level.
Entities
Institutions
- arXiv