- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
RAFT reduces training costs for memory-based temporal graph neural networks by using node-memory stability to filter prediction targets and sampled historical interactions. A loss-based controller adjusts when filtering is enabled. Experiments across three models and five datasets report 35.23% fewer inputs and 17.80% lower recorded training time on average, with small changes in validation performance. The contribution is a lightweight filtering mechanism that complements existing training systems, though its effectiveness depends on whether memory stability reliably indicates redundancy.
Critical strengths
-
The filtering mechanism avoids work early in the training pipeline. RAFT uses information from already executed memory updates, so selecting inputs does not require an extra prediction pass. Filtering before downstream preparation and computation makes the approach practically useful and complementary to existing training optimizations.
-
The experiments demonstrate meaningful savings across several settings. Evaluation covers three models, five datasets, and two training frameworks, with generally small changes in validation performance. The stage-level profiling also shows where the savings come from, providing evidence of actual computational benefits beyond simply processing fewer inputs.
Critical weaknesses
-
Filtering may miss changes involving previously stable nodes. RAFT uses past memory updates to decide whether to process a new event. If both endpoints were previously marked as stable, RAFT may skip a new event even when it reflects a change in interaction patterns. The loss-based controller may not detect this promptly, since it only observes losses on retained events.
-
Memory stability is not sufficient evidence that historical context is redundant. An interaction can remain informative even when its source node’s memory changes little; stable long-term preferences are one example. The reported loss curves show that filtering can work, but do not establish that stability identifies dispensable context.
Questions for authors
-
How can nodes marked as stable resume memory updates if their subsequent events keep being filtered?
-
At the same removal rate, how does filtering stable-source historical events compare with random filtering or filtering unstable-source events? This comparison would help establish whether memory stability identifies redundant context beyond the benefits of simply using less history.