Models that predict links in evolving networks, such as who will message whom or which user will review which product next, have looked far better than they are. They were graded against random node pairs that never interact, which a model can reject just by memorizing the graph's static shape. When the test instead uses historical negatives, meaning pairs that really did interact in the past but are inactive now, deep models lose sometimes over 20% AU-ROC. A lookup table of past edges then ranks second among all methods (Towards Better Evaluation for Dynamic Link Prediction). The deep models that do recover under harder tests are also slow. The Temporal Graph Benchmark reports inference up to two orders of magnitude slower than that lookup table. Meanwhile, spiking neural networks process information as sparse binary pulses, replacing multiply-accumulates with additions (0.9 pJ versus 4.6 pJ in 45nm CMOS, as cited in SpikingBERT). They have been applied to evolving graphs only for node classification, and to link prediction only on static graphs with random negatives.
We propose to build and test a spiking link predictor in which each node's state is a set of leaky integrate-and-fire neurons. Each neuron is a unit that sums its inputs, lets that sum decay over time, and fires a pulse only when the sum crosses a threshold. Interaction events drive these neurons, and the learned decay rates act as clocks for how long a relationship stays "alive." The model would be trained and graded on historical negatives. The claim we test is narrow: on the benchmarks where memorization fails, the leak gives a spiking model the right inductive bias to separate dormant from active pairs, and it does so at a small fraction of the arithmetic cost of the memory-based networks that currently win. Just as important is the honest negative outcome. If a hand-tuned decaying lookup table matches the spiking model, the "neuromorphic" story reduces to exponential smoothing, and we will report that.
Core Claim
A historical negative differs from a true positive mainly in how long ago the pair last interacted, and in whether the pair's neighborhoods are still active. A leaky neuron's membrane potential is a decaying trace of past input. With per-channel learnable time constants, a population of such neurons holds traces at several timescales, and it consumes compute only when events arrive. We claim that such a model has three properties. First, its accuracy drops much less than TGN's when the negatives switch from random to historical. Second, it beats a tuned, decaying edge lookup on high-surprise data, meaning data where most test edges were never seen in training. Third, it does this with an estimated synaptic-operation energy at least an order of magnitude below TGN at comparable ranking quality. No paper in our file tests any spiking model against historical negatives. Every spiking-graph result we found is therefore open to the critique that it learns structure rather than timing.
How This Differs From the Closest Work
Two spiking models already handle evolving graphs. Delay-DSGN learns a Gaussian synaptic delay per edge. ChronoSpike pairs adaptive spiking neurons, with learnable time constants and thresholds, with a small temporal Transformer, reaching 83–88% sparsity. Both motivate our design: Delay-DSGN's learned delays tracked how fast information moves in each dataset, which suggests learned time constants can encode edge lifespans. However, both models work on discrete snapshots, predict node labels, and never sample negative edges.
Spiking Variational Graph Auto-Encoders supplies the scoring method we reuse. It computes a weighted inner product of binary codes, scoring a pair by counting coincident spikes with a single learned weight vector, and so does link prediction with additions only (92.6% AUC on Cora with 20–80x lower operation energy). It works only on static graphs and only with random negatives. Our model is therefore not "a spiking graph network on a new task." It is the first spiking model whose state is driven by a continuous event stream, scored by addition-only coincidence, and trained and graded against dormant past edges.
Research Program
The model processes the timestamped edge stream in order. When an interaction $(u, v, t)$ arrives, it injects a current into the neuron populations of $u$ and $v$. That current comes from a small learned projection of the partner's most recent spike pattern and of edge features. Between events, each channel's membrane potential decays analytically according to the elapsed time:
$$V_c(t) = V_c(t_{\text{last}})\, e^{-(t - t_{\text{last}})/\tau_c},$$
where $V_c$ is channel $c$'s membrane potential, $t_{\text{last}}$ is that node's previous event time, and $\tau_c$ is a learnable time constant. Because of this, no compute is spent on idle nodes. Firing thresholds are learnable per channel, following ChronoSpike and Dynamic Reactive Spiking Graph Neural Network.
To score a candidate pair at query time $t$, the model decays both nodes' potentials to $t$, reads out their binary spike vectors $s_u, s_v \in \{0,1\}^d$, and computes $\sum_c w_c\, s_{u,c} s_{v,c}$. This is a weighted count of channels where both nodes fire, computed with additions only. A dormant pair's traces have decayed below threshold in the short-timescale channels, so those channels go silent. The model then has a direct way to separate "talked last week" from "talking now."
Training uses a ranking loss over negatives mixed at a fixed ratio of random to historical. We deliberately do not use the full multi-component curriculum of Curriculum Negative Mining, because the file flags its tuning overhead. A CurNM-trained variant appears as a comparison instead. Gradients through the spike step use a standard surrogate derivative. For long streams, we first use truncated backpropagation over event windows. As a second stage, we test whether the forward-in-time, constant-memory updates of NDOT extend to this per-node event setting. NDOT has only been shown on vision streams, so we treat the extension as a risk, not a premise.
Energy is reported as counted synaptic operations converted with the AC/MAC figures above, alongside measured CPU latency per event. The file notes that GPU simulation hides spike sparsity, so we will not claim measured hardware savings unless we obtain neuromorphic hardware access.
Baselines
The naive reference point is EdgeBank from Towards Better Evaluation for Dynamic Link Prediction. It is a training-free table that predicts "yes" for any pair seen before, in both its unlimited and time-window forms. We run it with the same TGB evaluator and the same negative sets. It matters because it sets the floor any learned model must clear under historical negatives, where it ranked second overall.
The most dangerous simple competitor is one we add ourselves: a decayed EdgeBank. It keeps, for each pair and each node, an exponentially decayed interaction count, with half-lives grid-searched on validation using the same tuning budget we give our model, and scores pairs by a tuned combination of those counts. This is our model with the learning removed. To claim the learned spiking dynamics matter, the spiking model must beat it on mean reciprocal rank (MRR), and especially on the high-surprise tgbl-review and on inductive negatives.
The strongest learned competitors are TGN, both plain and with CurNM, and NAT, run from the TGB reference implementations with equal hyperparameter search budgets. On tgbl-wiki, NAT's reported MRR is 0.749 (Temporal Graph Benchmark). On tgbl-review, TGB reports that GraphMixer and TGAT lead while NAT drops, but we found no reported numeric MRR on Lacuna for those models there. We will therefore rerun them. On the EdgeBank suite, CurNM's reported gain is an average 8.7% over TGN across 12 datasets, but the file gives no per-dataset historical-negative values. Beating these baselines means either matching their MRR within a small margin at roughly 10x fewer estimated operation-energy units, or exceeding them outright on any dataset.
Three ablations isolate the key ingredients. The first is an identical architecture with continuous activations, which separates the cost of spiking from its bias. The second trains the spiking model with random negatives only, which isolates historical-negative training. The third fixes the time constants instead of learning them, which tests whether the leak's learned timescales carry the signal.
Why not just…?
Why not just use a decayed EdgeBank? This is the real threat. If edge recurrence is all that matters, a tuned exponential decay on pair counts captures dormancy cheaply on any CPU. Evidence that it falls short exists only indirectly: memorization methods fail on high-surprise datasets like tgbl-review (Temporal Graph Benchmark) and cannot handle unseen nodes. Our spiking model scores pairs through learned neighborhood-driven spike patterns rather than pair identity, so it can in principle generalize. That claim is settled only by the head-to-head comparison. If the decayed table ties on tgbl-review and inductive negatives, we will conclude the spiking model adds energy-cheap smoothing but no reasoning.
Why not just compress TGN by quantizing or pruning it? A low-precision TGN keeps its gated memory, which Exploring Time Granularity found most robust under historical negatives. Even so, it updates dense GRU states at every event and still multiplies. Our continuous-activation ablation, combined with an 8-bit TGN run, measures whether quantization alone reaches the same accuracy-per-operation curve. If it does, the spiking design is not justified.
Why not just bolt a link-scoring head onto ChronoSpike? That model runs on discrete snapshots with a Transformer over time. It would need to choose a snapshot granularity, which the time-granularity study shows trades noise against resolution, and it would recompute every node at every snapshot. We will run this adapted ChronoSpike as an extra baseline on the smaller datasets, where snapshotting is feasible. This will show whether event-driven decay matters beyond spiking alone.
Proposed Experiments
We evaluate on TGB's tgbl-wiki, tgbl-review, tgbl-coin, and tgbl-comment using MRR under TGB's streaming protocol. We also use the EdgeBank suite (Wikipedia, Reddit, MOOC, LastFM, Enron, Social Evo, UN Vote, Contact), reporting AU-ROC under random, historical, and inductive negatives.
Four checks decide the outcome. First, the drop in AU-ROC from random to historical negatives must be smaller for the spiking model than for TGN. Second, the spiking model must beat the decayed EdgeBank on tgbl-review and on inductive negatives. Third, at matched MRR, estimated synaptic-operation energy and CPU latency per event must fall at least tenfold below TGN, and firing rates must be reported. Fourth, the learned time constants must track each dataset's typical gap between repeat interactions. That would mirror Delay-DSGN's finding that learned delays matched information speed, and would be evidence that the leak really encodes edge lifespan rather than acting as generic regularization.
References
- Towards Better Evaluation for Dynamic Link Prediction
- Temporal Graph Benchmark for Machine Learning on Temporal Graphs
- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation
- Delay-DSGN: A Dynamic Spiking Graph Neural Network with Delay Mechanisms for Evolving Graph
- ChronoSpike: An Adaptive Spiking Graph Neural Network for Dynamic Graphs
- Spiking Variational Graph Auto-Encoders for Efficient Graph Representation Learning
- Dynamic Reactive Spiking Graph Neural Network
- Curriculum negative mining for temporal networks
- NDOT: Neuronal Dynamics-based Online Training for Spiking Neural Networks
- Exploring Time Granularity on Temporal Graphs for Dynamic Link Prediction in Real-world Networks