AdaTKG: Adaptive Memory for Temporal Knowledge Graph Reasoning

2026-05-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

AdaTKG is a novel method for Temporal Knowledge Graph (TKG) reasoning that introduces adaptive, per-entity representations, departing from the static entity representations common in existing TKG methods. It maintains a per-entity memory, updated with each observed interaction, which accumulates online and refines predictions as more interactions occur. This memory update is instantiated as a learnable exponential moving average (EMA) governed by a single shared scalar parameter, enabling AdaTKG to handle entities unseen during training (emerging entities). The method fuses this adaptive memory with a static entity embedding via a learnable, d-dimensional adaptive gate. Extensive experiments on four TKG benchmarks (ICEWS14, ICEWS18, ICEWS05-15, GDELT) demonstrate that AdaTKG consistently outperforms strong TKG baselines, including the state-of-the-art TransFIR, particularly on emerging and unknown entities, with only marginal computational overhead.

Key takeaway

For research scientists developing TKG reasoning systems, AdaTKG offers a robust approach to improve performance on emerging and unknown entities. You should consider integrating adaptive, per-entity memory mechanisms, such as a learnable EMA, into your models. This strategy enhances predictive accuracy and training efficiency, especially when dealing with dynamic real-world data where entities frequently appear without prior history, making your systems more inclusive for applications like financial monitoring or public-health surveillance.

Key insights

Adaptive, per-entity memory significantly enhances temporal knowledge graph reasoning for emerging entities.

Principles

• Entity representations should adapt with interactions.
• Memory updates can be entity-agnostic for inductivity.
• Adaptive gates fuse static and dynamic information.

Method

AdaTKG refines entity representations using a per-entity memory updated via a learnable exponential moving average (EMA) and fuses it with static embeddings using an adaptive gate, ensuring inductivity for unseen entities.

In practice

• Implement per-entity memory for dynamic entity modeling.
• Use a shared scalar for EMA decay for efficiency.
• Employ an adaptive gate for graceful cold-start behavior.

Topics

• Temporal Knowledge Graphs
• Adaptive Memory
• Entity Representation
• Inductive Reasoning
• Exponential Moving Average

Code references

• seunghan96/AdaTKG

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Unlocking the Working Memory of Large Language Models for Latent Reasoning — Large language models can be trained to use working memory for efficient, latent reasoning, decoupling internal computation from external communication.
• Evolving Beyond Snapshots: Harmonizing Structure and Sequence via Entity State Tuning for Temporal Knowledge Graph Forecasting — Persistent entity states are crucial for robust, long-horizon Temporal Knowledge Graph forecasting, overcoming "episodic amnesia."
• APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI — APEX-MEM enhances long-term conversational AI through structured memory, temporal grounding, and intelligent retrieval.
• ADAG: Automatically Describing Attribution Graphs — Automated circuit tracing and interpretation can reveal LLM internal computations and identify steerable features.
• DYNA : Dynamic Episodic Memory Networks for Augmenting Large Language Models with Temporal Knowledge Graphs in Continuous Learning — DYNA augments frozen LLMs with temporal knowledge graphs to mitigate forgetting and improve temporal recall without retraining.
• A Better, Cheaper RAG (Neuro-Sym Multi-hop Reasoning) — TGS RAG fuses vector and graph search to overcome semantic drift and broken reasoning paths, reducing costs and improving accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.