Signed graphs have edges marked positive (trust, friendship) or negative (distrust, enmity). Models that predict the sign of a missing edge are trained mostly on a few kinds of easy examples. Positive edges are often 80–90% of the data. Most edges also sit in triangles that obey structural balance, the social-psychology rule that "the enemy of my enemy is my friend." The rare negative edges, and edges in triangles that violate balance, add little to the gradient. We propose to treat these edge groups as separate tasks inside one model. Their loss weights would be adjusted online, the way multi-task learning (training one shared network on several objectives) adjusts weights so that no objective starves the others.
The question is whether poor performance on contentious, balance-violating edges comes from how optimization effort is divided, rather than from the data itself. The leading fix today assumes the data is the problem. SGA builds an augmentation pipeline and an easy-to-hard curriculum that delays unbalanced edges. If online loss weighting can match or beat that curriculum without touching the graph, it gives a cheaper and more transparent fix. If it cannot, that is evidence that unbalanced edges need data-side treatment.
Core Claim
We partition the training edges into four groups by sign (positive or negative) and by local balance status. An edge counts as balanced or unbalanced according to the majority of the training-graph triangles it closes. Edges that close no triangles form a separate "unknown" bucket that is folded into its sign group. Each group's link-sign loss is one task.
The weight of each group is set from two signals:
- Gradient share. The first signal is the group's gradient norm at the last shared attention layer. This follows GradNorm, which balances tasks by equalizing gradient norms and their training pace.
- Validation gate. A validation-slope gate throttles any group whose held-out loss starts rising. This follows CoBa, which down-weights tasks whose validation curves climb.
The gate is essential rather than decorative. Balance-violating edges are partly genuine and partly label noise. Purely training-loss-driven schemes are known to amplify noisy tasks, and the validation signal is what keeps the method from overfitting them.
The claim is specific. On signed attention models such as SNEA and on SGCN, this weighting should raise accuracy on negative and unbalanced edges with no loss in overall AUC. It should also match SGA's curriculum without SGA's three-stage augmentation.
Research Program
The work has three stages.
Diagnostic. We first test whether the premise holds at all. We train SNEA and SGCN as published on Bitcoin-Alpha, Bitcoin-OTC, Epinions, Slashdot and Wiki-RfA. Throughout training we log each edge group's share of gradient norm in the shared node-projection and attention parameters. The proposal assumes that negative-unbalanced edges receive a gradient share far below their importance to the error. If their share already roughly matches their contribution to validation error, the program stops there, and that negative result is itself reportable.
Method. We add the group weighting as a drop-in loss wrapper. Each model keeps its own structural-balance regularizer at its published weight, and only the link-sign loss is split into groups. We deliberately do not target the balance-loss versus classification-loss ratio. It is a single scalar that a short grid search handles well, so it is not where a dynamic method can earn its keep. The added cost is four per-group backward passes to one shared layer, comparable to the roughly 5% overhead GradNorm reports.
Weight trajectories. We record the learned weight curves to see whether they implement an emergent curriculum. That would look like unbalanced edges being suppressed early and amplified late, which would mirror SGA's hand-designed schedule.
Baselines
Do nothing. The naive reference is each backbone trained as published, with uniform weight on all edges. The strongest reported numbers in the file set the bar:
- Bitcoin-Alpha: SNEA AUC 0.816, versus SGCN at 0.801 (SNEA).
- Bitcoin-OTC: SGCN AUC 0.823 and F1 0.925 (SGCN).
- Epinions: AUC 0.9333 for SiGAT, which applies attention over 38 balance- and status-theory motif neighborhoods, versus SGCN at 0.8818.
- Slashdot: no reported result found on Lacuna for Slashdot.
We will reproduce these under our own splits rather than compare against the printed figures.
Strongest competitor: SGA. The strongest existing method on this exact question is SGA. It scores edges as easy or hard by their local balance degree and feeds them through a curriculum after augmenting the graph. It reports only relative gains, not absolute numbers: 33.3% in binary F1 for a GAT backbone on Bitcoin-Alpha, and up to 22.2% in AUC for SGCN on Wiki-RfA. For the comparison to be fair, SGA and our method run on the same backbones, splits and epoch budget. Each gets the same number of tuning trials, about 30, on validation AUC.
Ablations that isolate the key ingredients. Three ablations each remove one ingredient:
- Sign-only grouping, which tests whether balance status matters beyond sign.
- No validation gate, which is pure gradient-norm balancing.
- Frozen average weights. Each group's dynamic weights are averaged over the run and reused as constants. This tests whether the timing of the weights matters or only their average level.
What counts as winning. Beating the baselines means matching or exceeding SGA's macro-F1 and AUC. It also means a statistically clear gain in AUC on the negative-unbalanced group over both the uniform and frozen-weight variants, across five seeds.
Why not just…?
Why not just reweight negative edges by inverse frequency? This is the most serious threat. It is simply a static, sign-only version of our method. Our design includes it directly as a baseline, and the frozen-average-weights ablation is its strongest form. If either matches the dynamic method, we will report that dynamic weighting adds nothing on signed graphs. The case for dynamics rests on one expectation: difficulty should shift as balanced positive edges saturate early. The logged weight trajectories will show whether that shift actually happens.
Why not just scale each group's loss by its moving average? That is cheaper because it needs no gradient computations, and Lakkapragada et al. show it flattens loss-scale differences and reduces negative transfer. However, Wang et al. and AdaMT-Net both find that equal loss magnitudes can hide very unequal gradient magnitudes. In sparse signed graphs, a group's gradient also depends on how many neighbors its endpoints aggregate. We run moving-average weighting as a baseline. If it ties, the cheaper method is the answer.
Why not just use SGA's curriculum? SGA alters the graph topology. SGA itself shows that careless structural edits sharply damage signed-graph performance, so a misjudged balance score could inject bias. Our method leaves the graph untouched. Which approach works better is an empirical question settled by the head-to-head comparison. A combined run, with SGA augmentation plus our weighting, tests whether the two are complementary or redundant.
Proposed Experiments
Benchmarks and metrics. We use Bitcoin-Alpha, Bitcoin-OTC, Epinions, Slashdot and Wiki-RfA with link-sign prediction. Metrics are AUC, macro-F1 and binary F1. We also report AUC stratified by the four edge groups. Test edges are stratified post hoc using their true sign and training-graph triangles; group labels are never needed at inference.
Decisive checks. The proposal's case rests on three checks:
- The diagnostic must show gradient starvation of the negative-unbalanced group.
- The full method must beat the frozen-weight and inverse-frequency variants on that group without lowering overall AUC.
- The full method must match SGA at lower preprocessing cost.
Failure case. A further check targets the noise risk. We inject label noise into a fraction of negative edges. If the validation gate is doing its job, the full method should degrade less than the ungated version, which should over-amplify the corrupted unbalanced group. If gating gives no robustness benefit, the CoBa ingredient is dropped.
References
- SGA: A Graph Augmentation Method for Signed Graph Neural Networks
- GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks
- CoBa: Convergence Balancer for Multitask Finetuning of Large Language Models
- Learning Signed Network Embedding via Graph Attention
- Signed Graph Convolutional Networks
- Signed Graph Attention Networks
- Mitigating Negative Transfer in Multi-Task Learning with Exponential Moving Average Loss Weighting Strategies (Student Abstract)
- When and why PINNs fail to train: A neural tangent kernel perspective
- AdaMT-Net: An Adaptive Weight Learning Based Multi-Task Learning Model For Scene Understanding