Temporal Graph Learning Systems
Building scalable, staleness-free training infrastructure for memory-based temporal graph neural networks — from custom CUDA kernels to distributed multi-node systems.
The Problem
Temporal graph neural networks (TGNNs) process continuously evolving graphs where nodes maintain a memory state that accumulates information from past interactions. The challenge: when events in the same training batch depend on each other, standard parallel execution reads stale memory — memory that has not yet been updated by earlier events in the same batch. This staleness grows with batch size and degrades model accuracy, creating a fundamental tension between training throughput and model correctness.
Existing systems either process events fully sequentially (correct but slow) or ignore intra-batch dependencies entirely (fast but inaccurate). Neither scales. The goal of this project is to build systems that are both — staleness-free and GPU-efficient.
What We Built: PRISM
Memory Computation Graph (MCG)
We model intra-batch memory dependencies as a directed graph — the MCG — where nodes are events and edges encode which memory reads depend on which prior writes. Custom CUDA kernels traverse this graph to propagate fresh memory states, resolving staleness without serializing the whole batch.
Lazy Freshness
Rather than enforcing strict sequential ordering, PRISM uses lazy freshness: each event is guaranteed a memory state that reflects all events it causally depends on, but independent events still execute in parallel. This unlocks significant GPU parallelism that strict staleness-free methods forfeit.
GRN-Stream Sampler
A streaming temporal neighbor sampler that decouples data ingestion from training. Graph edges arrive continuously; the sampler maintains a chunk-based index (TCI) for efficient neighbor lookups without rebuilding the full index each step.
Model-Agnostic Design
PRISM is a training system wrapper, not a new model. TGN, TNCN, APAN, and other memory-based TGNNs plug in unchanged — PRISM handles all memory versioning and dependency resolution transparently at the system level.
Results
Accuracy
Up to 28% higher accuracy versus stale-memory baselines across five temporal graph benchmarks at the same batch sizes.
Leaderboard
Results up to 9.2% above the TGB leaderboard, achieved by applying PRISM to existing unmodified models.
Throughput
Training time remains competitive with parallel stale-memory baselines and is often lower than strictly sequential staleness-free execution.
Scalability
Accuracy improvements grow with batch size — the regime where prior systems degrade most. PRISM gets more useful as scale increases.
What’s Next
- Multi-node distributed training — extending the MCG and lazy freshness framework across multiple GPUs and machines with NCCL-based memory synchronization.
- Relational temporal graphs — applying this infrastructure to heterogeneous relational databases for predictive queries (see Predictive Query project).
- Provenance graph learning — reusing PRISM’s dynamic graph engine for OS-level intrusion detection over audit log streams (see Intrusion Detection project).
Paper & Resources
VLDB’26 Publication
Md Ashraful Islam, Hojae Son, Suhaas Kiran Doddagaddavalli Gangadharaiah, Marco Serafini
Proceedings of the VLDB Endowment, 2026
Citation
@article{islam2026prism,
title = {PRISM: A Training System to Unlock
the Potential of Temporal Graph
Learning Through Staleness Avoidance},
author = {Islam, Md Ashraful and Son, Hojae
and Kiran, Suhaas and Serafini, Marco},
journal = {Proc. VLDB Endow.},
year = {2026}
}