Graph meta-learning trains one model on many small labeling problems, called episodes, so that it can classify nodes of new classes or new graphs from one to five labeled examples. Every current method in this area quietly assumes that all episodes want the same kind of neighborhood averaging. That assumption fails on mixed collections of graphs. In a citation network, connected papers usually share a label, so averaging over neighbors helps. In a web-page or rating network, connected nodes often have different labels, so averaging erases the signal. A single shared encoder trained on both kinds of graph is pulled in opposite directions.
We propose that a large share of negative transfer in graph meta-learning is a frequency conflict of this kind. Negative transfer means the meta-learned model ends up worse on new tasks than a model trained without the other tasks. In a frequency conflict, tasks disagree on whether the encoder should smooth node features toward their neighbors or sharpen them away from neighbors. Our remedy is narrow. The encoder keeps separate low-pass (averaging), high-pass (difference-from-neighbors) and identity channels. Their feature weights are shared across all episodes. Only three per-task mixing scalars are adapted from the support set during the few-shot inner loop. The claim is falsifiable: if the conflict explanation is right, the penalty for meta-training on mixed graphs should grow with the homophily gap between source and target, and task-adapted mixing should remove most of that penalty. Homophily is the fraction of edges joining same-label nodes. Neither a fixed filterbank nor a rule that picks a filter from estimated homophily should match it.
Core Claim and Positioning
Four papers already join few-shot graph meta-learning to smoothing, and none adapts the frequency response per task.
The first two treat smoothing as a matter of locality. Graph Meta Learning via Local Subgraphs (G-META) encodes fixed $h$-hop subgraphs with prototypes and MAML, and argues this bounds global over-smoothing. Its authors name negative transfer as a vulnerability but offer no mechanism against it. Contrastive Meta-Learning for Few-shot Node Classification (COSMIC) uses Personalized PageRank subgraphs and similarity-sensitive mix-up. It reports sensitivity to subgraph size, because large subgraphs over-smooth locally. Both papers bound how far averaging reaches. Neither questions whether averaging is the right operation for a given task, and that question is the one we ask.
The other two treat smoothing as something to calibrate within a single regime. Task-Adaptive Few-shot Node Classification (TENT) adapts at the node, class and task levels. It does so partly by wiring each class's support nodes to a virtual hub, which deliberately smooths them together. That is a low-pass prior we expect to hurt on heterophilic tasks. Robust Graph Meta-Learning via Manifold Calibration with Proxy Subgraphs (Pro-MC) calibrates between a noise-rugged embedding and an over-denoised one, but only for edge noise within one graph family.
All four papers evaluate on largely homophilic citation and co-purchase graphs. Our proposal differs from all four in what gets adapted per episode: the sign and balance of the graph filter, not the subgraph radius, class anchors or edge set.
The mechanism comes from the over-smoothing literature. Complete the Missing Half splits aggregation into low-pass and high-pass channels, $I = L_{LP} + L_{HP}$, with learnable mixing scalars. It gains roughly 10% on Texas and Cornell over plain GCN. Geometric Scattering Attention Networks (GSAN) attends over spectral channels per node and reaches 61.2% on Chameleon versus 28.2% for GCN. Both papers learn the mixing once per dataset or per node, in ordinary supervised training. We move the mixing into the meta-learning inner loop, which makes it a per-task quantity inferred from a handful of labels.
The motivation for treating negative transfer as conflict between task optima comes from two sources. Meta Omnium documents a multi-task penalty: meta-training jointly on heterogeneous task families is worse than training on each family alone. It also finds that Prototypical Networks have the best average rank. Variational Continual Bayesian Meta-Learning shows that compressing dissimilar tasks into one initialization degrades all of them. That paper answers with a mixture of parameter sets. We argue that, for graphs, a three-dimensional task variable can do the same job as a parameter mixture.
Research Program
Each episode's encoder computes a node representation of the form
$$Z_\tau = \alpha^{\tau}_{L}\, L_{LP} X W_L + \alpha^{\tau}_{H}\, L_{HP} X W_H + \alpha^{\tau}_{I}\, X W_I .$$
Here $X$ is the node feature matrix of the episode's subgraph and $L_{LP}$ is the normalized averaging operator. The high-pass operator is $L_{HP} = I - L_{LP}$. The matrices $W_L, W_H, W_I$ are channel weights shared across all episodes, and $\alpha^\tau = (\alpha^\tau_L, \alpha^\tau_H, \alpha^\tau_I)$ are the task's mixing scalars. We stack two such layers. Classification is by distance to class prototypes, following G-META and the Meta Omnium finding that prototype-based learners are most robust.
The inner loop takes a few gradient steps on $\alpha^\tau$ alone, using the support-set prototype loss and starting from a meta-learned initialization $\alpha^0$. The outer loop updates $W_L, W_H, W_I$ and $\alpha^0$ on the query loss. Because only three numbers change per task, the inner loop is cheap and hard to overfit with one shot per class. It cannot express any task difference other than filter shape. That restriction is intentional, because it makes the hypothesis testable.
The analysis component measures two things. The first is the multi-task penalty, the target-task accuracy of a model meta-trained only on graphs of the target's homophily type minus that of a model meta-trained on the mixed pool. The second is the effective rank of query embeddings, as defined in Are we measuring oversmoothing in graph neural networks correctly?. That paper shows effective rank tracks accuracy with average Pearson correlation 0.91 where Dirichlet energy fails, and it cautions that high rank does not imply useful features. We therefore use rank only as a one-sided diagnostic: a drop in rank signals collapse, but a high rank is not counted as success.
The specific prediction is about heterophilic target episodes. On those episodes, a low-pass-only meta-learner trained on the mixed pool should show both a positive penalty and depressed effective rank relative to its homophily-matched counterpart, and task-adapted mixing should restore both.
Proposed Experiments
Benchmark. We construct a mixed-homophily episodic benchmark from graphs in the file. The homophilic graphs are CoraFull, DBLP, Amazon-E and ogbn-arxiv. The heterophilic graphs are Chameleon (homophily 0.23), Squirrel, Amazon Ratings, Texas and Cornell.
The small heterophilic graphs have only a handful of classes, so a novel-class split inside them is not possible. We therefore use G-META's multiple-graph setting with leave-one-graph-out evaluation: each held-out graph is the meta-test source, with $N$ capped at its class count and $K \in \{1, 3, 5\}$. Feature spaces differ across graphs, so every graph's features are projected to a common dimension by truncated SVD before any method sees them. This projection is itself a source of transfer loss. It is applied identically to every method, and we report it as a limitation.
Single-graph sanity check. On the standard single-graph novel-class splits we reproduce the existing reference numbers. COSMIC reports 74.93% on CoraFull 5-way 1-shot against roughly 60.44% for G-META, per Contrastive Meta-Learning for Few-shot Node Classification. TENT reports 82.84% on DBLP 5-way 5-shot against 80.85% for GPN, per Task-Adaptive Few-shot Node Classification. No reported few-shot result was found on Lacuna for any method on Chameleon, Squirrel, Texas, Cornell or Amazon Ratings. The GSAN number on Chameleon comes from fully supervised training and is not a few-shot reference.
Metrics. We report mean accuracy with 95% intervals over 600 test episodes, the multi-task penalty per held-out graph, and query effective rank.
Baselines
The naive reference point is a two-layer GCN encoder with prototype classification, meta-trained on the pooled mixed graphs. Its homophily-matched twin is the same model meta-trained only on graphs of the target's type. The gap between these two defines the penalty we aim to close. This oracle uses information about the target type that our method does not get, so matching it counts as complete success.
The strongest existing methods are COSMIC and TENT. Both report the best numbers in the file on the few-shot node classification benchmarks, and both use low-pass smoothing priors (Personalized PageRank diffusion and virtual class hubs) that we predict will break on heterophilic targets. G-META is included as the multiple-graph reference, since it was designed for exactly that setting. Each method is run from its released settings. Each then receives the same hyperparameter search budget, which we fix at 30 configurations per method, selected on held-out validation graphs. All methods get the same SVD features, the same episode sampler and the same GPU-hour cap. Pro-MC is omitted from the main table because its optimal-transport loop targets edge noise, not task mixing. It is run only on its own homophilic graphs as a check.
Two ablations isolate the key ingredient. The first, a fixed filterbank, is the identical three-channel encoder with $\alpha$ meta-learned once and shared across tasks, which amounts to FB-GNN inside a prototype meta-learner. The second, GSAN-style per-node channel attention without task conditioning, tests whether per-node adaptivity already suffices.
Beating the baselines means two things. Across the heterophilic held-out graphs, our method must close at least half the penalty, measured as an average improvement over the best of COSMIC, TENT and G-META with non-overlapping intervals. It must also stay within one point of COSMIC on homophilic held-out graphs. Most importantly, it must beat the fixed-filterbank ablation. If it does not, the task-adapted part of the idea has failed, even if the filterbank itself helps.
Why not just…?
Why not just keep the encoder shallow? Two layers already bound smoothing, but the Chameleon result in Geometric Scattering Attention Networks shows that GCN collapses to 28.2% there even at shallow depth. On heterophilic graphs the harm comes from the direction of averaging, not its depth. Our encoder is itself two layers, so this baseline differs from ours only in filter shape. The naive reference already tests it.
Why not just use a fixed filterbank or node-wise spectral attention, with no task adaptation? This is the real threat. A per-node attention model might learn to recognize heterophilic neighborhoods from features alone. If the fixed-filterbank or GSAN-style ablation matches our method on held-out graphs, we will conclude that task-level adaptation is unnecessary. The publishable finding would then be the diagnostic one: frequency conflict explains the penalty, and a static filterbank fixes it. We also check whether the learned $\alpha^\tau$ correlates with true episode homophily. A strong correlation alongside equal accuracy would mean the per-node model is implicitly doing the same inference.
Why not just estimate homophily from the support labels and pick a filter by rule? With one to five labels per class, support-only homophily estimates on small subgraphs are noisy. Edges among support nodes are also rare. We run this rule-based variant directly: it sets $\alpha$ by thresholding estimated support-subgraph homophily. Matching it would show that gradient adaptation adds nothing beyond a cheap estimator, and we would report that outcome as such.
Why not just pre-train on the pooled graphs and fit a linear probe? Meta Omnium finds that pooled pre-training does not remove cross-family conflict. Few-shot initializing of Active Learner via Meta-Learning shows that irrelevant source tasks hurt initializations. Both results come from settings other than graphs, so we include a pooled-pretraining-plus-probe baseline under the same budget rather than assume the result transfers.
Decisive Checks
The proposal stands or falls on three outcomes.
First, the penalty for the naive pooled model must rise with the homophily gap between source pool and target, measured across held-out graphs. If it does not, frequency conflict is not the dominant cause of negative transfer, and the method has no rationale.
Second, the adapted $\alpha^\tau_H / \alpha^\tau_L$ ratio must track true episode homophily on held-out graphs without ever having seen it.
Third, our method must beat the fixed-filterbank and rule-based ablations on heterophilic targets while staying within one point of COSMIC on homophilic ones. Collapsed query rank on heterophilic targets for the low-pass baselines, recovering under our method, would support the mechanism. Following the caveat in Are we measuring oversmoothing in graph neural networks correctly?, rank recovery without accuracy recovery counts as a failure.
References
- Graph Meta Learning via Local Subgraphs
- Contrastive Meta-Learning for Few-shot Node Classification
- Task-Adaptive Few-shot Node Classification
- Robust Graph Meta-Learning via Manifold Calibration with Proxy Subgraphs
- Complete the Missing Half: Augmenting Aggregation Filtering with Diversification for Graph Convolutional Networks
- Geometric Scattering Attention Networks
- Meta Omnium: A Benchmark for General-Purpose Learning-to-Learn
- Variational Continual Bayesian Meta-Learning
- Are we measuring oversmoothing in graph neural networks correctly?
- Few-shot initializing of Active Learner via Meta-Learning