Simple temporal link predictors now beat elaborate ones. A one-layer mixing network called GraphMixer beats recurrent and attention-based models on interaction logs. A memoryless cross-attention model called CRAFT tops 10 of 17 temporal benchmarks. The usual reading is that these models learn temporal dynamics more efficiently. A cheaper explanation fits the same results. Both models see how many recent interactions a node has had. Nodes that were active recently tend to be active again, the "rich get richer" process known as preferential attachment. A model could score well simply by learning to bet on popular nodes. Nobody has separated these two explanations.
We propose a way to test, component by component, how much of a trained model's accuracy comes from cumulative advantage. The method is an ablation in which a component's input is replaced by a randomized version of the graph. This version keeps every node's interaction counts and timing but scrambles who interacted with whom. If accuracy survives the scramble, the component was exploiting popularity. If accuracy collapses, it was using real relational information. Because the scrambled inputs keep the same degree statistics, they stay close to the training distribution. That avoids the out-of-distribution shock that zeroing or deleting a module causes.
Core Claim
Our hypothesis is that a large, measurable share of GraphMixer's and CRAFT's advantage over older temporal models is preferential attachment. We expect this share to sit in specific parts of each model: GraphMixer's zero-padded link encoder, whose padding length reveals how much history a node has, and CRAFT's learnable node embeddings. The new contribution is not the observation that link predictors are biased toward high-degree nodes. It is a component-level attribution that uses degree-preserving null models as the reference distribution for ablation, and the per-module numbers that attribution produces.
Relation to Prior Work
The motivating evidence comes from the preferential attachment literature. Measuring preferential attachment in evolving networks measured link acquisition as a function of degree from time-stamped histories. Experience versus talent shapes the structure of the Web showed that accumulated popularity amplifies small quality differences into large visibility gaps. Disentangling the two is a known open problem, and our scrambles are designed to separate them inside a learned model.
Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction proves that symmetric-normalized graph convolutional networks score links roughly in proportion to $\sqrt{\deg(i)\deg(j)}$. It reports a Pearson correlation above 0.8 between that quantity and trained scores, and it mitigates the bias with a training-time regularizer. That work is static, covers only convolutional networks, and is correlational at the level of the whole model. We ask the causal, per-module version of its question for temporal architectures, where the proof does not apply.
The deconstruction papers are our objects of study rather than competitors. Do We Really Need Complicated Model Architectures For Temporal Networks? introduced GraphMixer and reports 96.31% average precision on LastFM against 82.23% for TGN. Future Link Prediction Without Memory or Aggregation introduced CRAFT and reports a 49% relative gain on ML-20M. Each removed modules and found that accuracy held up. Neither asked what signal the remaining modules carry, and that is the question we pursue.
The methodological lineage is ablation practice. Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small replaces activations with a reference distribution instead of zeros, a technique called mean ablation. We borrow that idea and choose a structural null model as the reference. Two critiques motivate the rigor. Troubling Trends in Machine Learning Scholarship warns against crediting architecture for gains that come from elsewhere. Blindfold Baselines for Embodied QA showed that agents ignored vision and relied on language priors; we look for the graph analogue, a model that ignores relational structure and relies on degree.
Can GNNs Learn Link Heuristics? constrains the design. It shows that pooling operators wash out explicit counting signals, so we cannot assume a model computes a degree feature just because the information is present. Our method measures whether the model uses degree rather than whether it could.
Research Program
We define three input-level null models over the temporal event stream visible to a module.
The identity scramble permutes partner identities across events. It preserves each node's event count and timestamps, and therefore its full degree trajectory, while destroying who-links-to-whom. This scramble approximates "popularity only."
The order scramble keeps each node's multiset of partners but permutes event times within the window. This isolates fine-grained temporal order.
The uniform scramble replaces partners and counts with random draws. It destroys degree information too and serves as the control for the key ingredient.
Each scramble is applied at inference to one module's input at a time. For GraphMixer those modules are the link encoder, the node encoder that mean-pools over neighbors, and the time encoding. For CRAFT they are the neighbor sequence fed to cross-attention and the learnable node embeddings, which we replace with embeddings of degree-matched other nodes. We also run all pairwise combinations, because single-module removal misses interactions.
The headline quantity is the cumulative-advantage share of a module $m$:
$$ S_m = \frac{\mathrm{AP}_{\text{id-scr}(m)} - \mathrm{AP}_{\text{unif-scr}(m)}}{\mathrm{AP}_{\text{full}} - \mathrm{AP}_{\text{unif-scr}(m)}} $$
Here $\mathrm{AP}$ is average precision. $\mathrm{AP}_{\text{full}}$ is the intact model. $\mathrm{AP}_{\text{id-scr}(m)}$ is the score with only module $m$'s input identity-scrambled, and $\mathrm{AP}_{\text{unif-scr}(m)}$ is the score with that input uniformly scrambled. $S_m$ near 1 means the module's contribution is fully explained by degree trajectories. $S_m$ near 0 means it depends on who the partners are.
Inference-time scrambling can understate what a model would learn under different data. We therefore also retrain each model on identity-scrambled training streams and report the retrained accuracy alongside the inference-time result.
Baselines
The naive reference point is a zero-parameter temporal popularity score: the destination's interaction count within the same window $T$ that GraphMixer uses, multiplied by the source's count. This is temporal preferential attachment written as a heuristic. It is run on identical splits and negatives with no tuning beyond choosing $T$ from GraphMixer's own grid. There is no reported result found on Lacuna for this heuristic on LastFM, Wikipedia, Reddit, MOOC, or GDELT, so we compute it ourselves.
The strongest existing methods are GraphMixer and CRAFT themselves. The reported figures are 96.31% average precision on LastFM for GraphMixer (Do We Really Need Complicated Model Architectures For Temporal Networks?) and a 49% relative gain on ML-20M for CRAFT (Future Link Prediction Without Memory or Aggregation). We retrain both with the authors' hyperparameters and check that we roughly reproduce those numbers before any ablation, so that attribution is done on faithful models.
The competing attribution baselines are standard zero ablation of each module and a correlational analysis in the style of Networked Inequality, which correlates model scores with the log degree-product. Both receive the same trained checkpoints.
The ablation that isolates our key ingredient is identity scramble versus uniform scramble. If degree preservation does not change the measured contribution, our null-model choice adds nothing over ordinary randomization, and we will report that.
Our method beats the attribution baselines if it makes correct predictions where they fail. Its per-module $S_m$ should predict how each module's accuracy changes under degree-matched negative sampling, while zero-ablation drops and score–degree correlations should predict that change less well. This is checked across datasets by rank correlation.
Why not just…?
Why not just run the popularity heuristic and stop? If the heuristic alone matched GraphMixer on LastFM, the headline question would be answered without any ablation, and we treat this as a real threat. The heuristic is our first experiment, and a near-tie would itself be a finding worth reporting. It cannot tell which module carries the signal, though. It also cannot tell whether a model combines popularity with relational information in a way the heuristic misses. A model can beat the heuristic and still rest mostly on degree. The retrained identity-scramble runs settle how much of the gap above the heuristic is non-popularity signal.
Why not just zero out each module? Zeroing GraphMixer's link encoder produces an all-padding input the model never saw in training. An accuracy drop then mixes "this module mattered" with "this input broke the model." The research file lists this out-of-distribution problem as a known limitation of ablation. The identity scramble keeps padding length and timing statistics intact. The two methods can be compared directly by seeing which better predicts the degree-matched evaluation results.
Why not just evaluate with degree-matched negatives? Harder negatives are valuable, and we use them as ground truth for validating the attribution. A benchmark change alone reports that accuracy fell. It does not say which component was responsible or whether retraining would recover the loss. It also changes the task, whereas our attribution explains performance on the task as published.
Proposed Experiments
The primary suite is LastFM, Wikipedia, Reddit, MOOC, and GDELT, as evaluated in GraphMixer's paper, plus ML-20M and the TGB-Seq datasets used by CRAFT, with WikiLink as a scale test. For every dataset we report: the popularity heuristic; the full models; per-module and pairwise $S_m$ under inference-time scrambling; accuracy after retraining on identity-scrambled and order-scrambled streams; and accuracy under negatives matched to the true destination's degree. As a static sanity check, we apply the identity scramble to a graph convolutional network on DBLP-Fairness and ogbl-collab. There it should reproduce the degree dominance that Subramonian et al. predict, which validates the instrument on a case with a known answer.
Three outcomes would be decisive. First, GraphMixer's link encoder should have $S_m$ above 0.7 on LastFM, and its advantage over TGN should largely vanish under degree-matched negatives. That would show the reported 14-point gap is mostly cumulative advantage. Second, CRAFT's cross-attention should show a markedly lower $S_m$ than its node embeddings. That would locate CRAFT's popularity reliance in its transductive identifiers rather than its attention. Third, if instead every module's $S_m$ stays low and the order scramble causes the largest drops, the minimalist models really are learning temporal structure, and the cheap explanation is refuted. Any of these outcomes changes how future temporal architectures should be credited.
References
- Measuring preferential attachment in evolving networks
- Experience versus talent shapes the structure of the Web
- Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction
- Do We Really Need Complicated Model Architectures For Temporal Networks?
- Future Link Prediction Without Memory or Aggregation
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small
- Troubling Trends in Machine Learning Scholarship
- Blindfold Baselines for Embodied QA
- Can GNNs Learn Link Heuristics? a Concise Review and Evaluation of Link Prediction Methods