Not All Memories Age the Same: Autodiscovery of Adaptive Decay in Knowledge Graphs

arXiv cs.LG / 5/1/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that knowledge graphs treat all facts as equally current, and that using a single uniform decay/forgetting curve is fundamentally mispecified for temporal retrieval.
It introduces a hierarchical adaptive decay framework that models continuous decay using two orthogonal signals: velocity (observation frequency) and volatility (how much values change, estimated via embedding distance).
The model decomposes decay into three learnable levels—domain-level, context-level, and entity-level—so different predicates and contexts can have inherently different aging behaviors without requiring predefined taxonomies.
By casting edge lifetime as a survival problem based on value supersession rather than simple re-observation, the approach better identifies what should be considered important at query time.
Experiments show strong recovery on synthetic graphs (HDBSCAN ARI = 1.0) and that, on Wikipedia and Synthea clinical EHR data, velocity–volatility clusters follow Lindy-like behavior while uniform decay is 18x worse than no temporal weighting.

Abstract

Knowledge graphs used for retrieval treat all facts as equally current. Existing temporal approaches apply uniform decay, using a single forgetting curve regardless of knowledge type. We show this is fundamentally misspecified: different knowledge types exhibit different temporal dynamics, and the core retrieval problem is not latency or throughput but identifying what is important at query time. We propose a hierarchical framework that replaces uniform decay with a continuous decay surface parameterized by two orthogonal signals: velocity (how frequently a concept is observed) and volatility (how much the value changes between observations, measured via embedding distance). The decay surface is decomposed into three learnable levels: domain-level parameters capture universal patterns (some predicates are inherently permanent, others inherently transient), context-level parameters capture setting-dependent variation, and entity-level adaptation personalizes decay to specific subjects. All parameters emerge from data through survival analysis on observed value lifetimes, requiring no predefined taxonomies or domain expertise. We formulate edge lifetime as a survival problem where the event is value supersession (a meaningfully different value replacing the current one), distinct from mere re-observation. Experiments on synthetic temporal knowledge graphs demonstrate recovery of planted hierarchical parameters (HDBSCAN ARI = 1.0). Validation on 107 Wikipedia articles and 1,163 patient records from the Synthea clinical EHR simulator shows that velocity-volatility clusters emerge naturally, align with observable persistence patterns, and near-universally exhibit the Lindy effect (Weibull shape k < 1). Uniform decay performs 18x worse than no temporal weighting. Heterogeneous decay recovers from this, with each hierarchy level contributing measurable improvement.