Research map / Graph learning / Graph self-supervised learning

Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting

NeurIPS 2026 · Graph learning › Graph self-supervised learning · approach: Semi-supervised node representation learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Spectral Reverse Prompt (SRP), a graph prompt-tuning method that reweights singular directions of a frozen GNN encoder with a learnable soft-threshold mask and adds a null-space PCA channel, reporting gains over graph prompting baselines on node and graph classification.

Problem

Parameter-efficient adaptation of frozen self-supervised GNN encoders for few-shot node and graph classification, where prompting methods ignore the encoder's internal spectral structure and dominant singular directions from pre-training may misalign with downstream tasks.

Approach

Treats the frozen encoder's weight SVD as a spectral resource: a sigmoid soft-threshold mask down-weights dominant singular directions and amplifies weak ones, while a null-space channel adds PCA features from directions the encoder cannot express, with both channels fused into layer-wise prompts.

Key results

• On 5-shot node classification with GraphCL pre-training, SRP reaches 68.11% on Cora and 49.02% on CiteSeer, versus 63.35% and 45.98% for the strongest baseline GraphLoRA.
• On graph classification with GraphCL pre-training, SRP reaches 69.71% on DD and 69.35% on Mutagenicity, the best among compared methods in those columns.
• Ablations show removing the null-space channel costs about 6.3 points on average (GraphCL) and removing reversal (uniform weighting) costs about 2.2 points.

Surprising findings

• Weak-k selection (keeping the weakest singular directions) outperforms strong-k and random-k selection of matched cardinality, supporting the spectral reversal claim.
• On graph-level datasets with small input dimension, the projective fallback is used, and gains are smaller on those tasks.

Limitations

• Spectral reweighting uses a shared singular basis and a single global threshold, so it may miss region-specific frequency heterogeneity.
• Null-space channel effectiveness depends on input dimension versus encoder rank and is limited when the frozen transformation has little or no null space.
• Many reported results have large standard deviations, and some baselines differ from the reported numbers in other settings; the paper's theoretical bounds are loose (O-notation with a sketch-level proof).

Experimental setup

Datasets: Cora CiteSeer PubMed Flickr ogbn-arxiv ENZYMES DD NCI1 NCI109 Mutagenicity Cornell Squirrel

Benchmarks: Cora CiteSeer PubMed ogbn-arxiv Flickr ENZYMES DD NCI1 NCI109 Mutagenicity Cornell Squirrel

Metrics: accuracy (%) mean ± std over 5 seeds

Baselines: Classifier Only GPPT GraphPrompt All-in-one GPF GPF-plus EdgePrompt EdgePrompt+ GraphLoRA

Builds on: GraphCL SimGRACE GPF EdgePrompt GraphLoRA Prog

Closest papers by meaning

• GraphTOP: Graph Topology-Oriented Prompting for Graph Neural Networks (NeurIPS 2025 · similarity 0.89)
• One Prompt Fits All: Universal Graph Adaptation for Pretrained Models (NeurIPS 2025 · similarity 0.89)
• Asymmetric Learning for Spectral Graph Neural Networks (AAAI 2025 · similarity 0.89)
• Edge Prompt Tuning for Graph Neural Networks (ICLR 2025 · similarity 0.89)
• Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025 · similarity 0.88)
• HeterGP: Bridging Heterogeneity in Graph Neural Networks with Multi-View Prompting (AAAI 2025 · similarity 0.88)