Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors

AI summary

Overview

• Research area: Graph representation learning / graph neural networks, specifically multi-task and cross-task transfer learning between node classification (NC) and link prediction (LP) on a single shared graph.
• Technical level: Advanced. The paper is written primarily as a benchmarking and protocol paper, but it assumes familiarity with GNN backbones, transductive/inductive splits, negative sampling for link prediction, and metrics such as ROC-AUC and accuracy.
• Scope (one sentence): The paper formalizes same-graph NC–LP transfer, proposes a leakage-free evaluation protocol, evaluates five transfer mechanisms across three GNN backbones on 11 datasets in both transfer directions, and identifies simple diagnostics (especially homophily) that predict when transfer helps or hurts.

What This Paper Is About

Most graph learning work trains and evaluates a GNN for a single objective — either node classification or link prediction — even though real graphs usually carry both node labels and observed edges at the same time. The paper asks a practical question: when both supervision signals coexist on the same graph, can supervision from one task improve the other, and in which direction? Its goal is not a new architecture but a clean, leakage-free evaluation protocol that makes this bidirectional transfer measurable and predictable.

Key Contributions

1. Problem setting — same-graph cross-task learning. The authors formalize NC and LP reuse on a single underlying graph with coexisting supervision signals, and study both directions (NC to LP and LP to NC) rather than a single objective.
2. Leakage-free evaluation protocol. A standardized setup that fixes node splits, edge splits, the observed message-passing graph, and LP negatives across all methods, so that evaluated edges cannot leak into embeddings through message passing and comparisons are not confounded by re-sampled negatives.
3. Systematic study across backbones and transfer strategies. Five lightweight cross-task transfer strategies (WS, ET-Rep, ET-Concat, MV, Joint) are evaluated across three backbones (GCN, GraphSAGE, GPS), documenting a directional asymmetry that persists across architectures and regimes.
4. Multi-objective selection metric and predictors. The CoTask Score (CTS) is introduced to summarize joint NC+LP utility when one training regime must serve both tasks, along with interpretable diagnostics: homophily predicts NC-to-LP gains, while LP-to-NC gains concentrate in structure-dominant regimes with high LP learnability and NC headroom.

Main Findings

• Transfer is strongly directional. NC to LP is consistently beneficial on homophilic graphs, while LP to NC is fragile and can degrade node classification accuracy under naive representation reuse.
• NC to LP is robust on homophilic graphs. On Cora, Citeseer, and PubMed, every backbone achieves consistent AUC gains from NC to LP across multiple mechanisms. For example, Cora's GCN LP baseline AUC is 80.3 with gains of 9.4 (WS), 10.7 (ET Rep), 9.8 (ET Concat), 10.1 (MV), and 12.3 (Joint).
• Mechanisms that couple tasks during training often win. MV and Joint — which explicitly combine node supervision with graph structure during training — frequently outperform warm-start style transfer, suggesting the benefit is representational rather than purely an optimization effect.
• Naive LP-to-NC reuse can be severely harmful. The ET-Rep mechanism produces large accuracy drops: −12.7 (Cora, GCN), −18.7 (Cora, GraphSAGE), −18.3 (Cora, GPS), −16.8 / −25.3 / −27.8 on Citeseer, −21.4 / −20.4 / −23.1 on PubMed, and −28.2 / −27.5 / −34.5 on Roman.
• A second, structure-dominant regime reverses the picture. LP to NC becomes beneficial when LP is easy but NC is unsaturated. Usa, Europe, and Brazil satisfy this criterion across all three backbones, showing positive LP-to-NC gains even where homophily is low — for example, Usa with GPS Joint at +20.5 and ET Rep at +24.2, Europe with GraphSAGE ET Concat at +17.9, and Brazil with GCN ET Rep at +15.2 and Joint at +13.7.
• The structure-dominant regime is defined reproducibly. Under a fixed backbone and protocol, a dataset is structure-dominant if its baseline LP AUC ranks in the upper half of the benchmark while its baseline NC accuracy ranks in the lower half — a threshold-free criterion given a fixed dataset collection.
• Backbone gains are not explained by model size. Joint uses the same encoder capacity as single-task baselines and only adds a lightweight second head; MV trains two task-specific encoders but each task is evaluated with its corresponding encoder (no ensembling or backbone expansion at test time).
• Simple statistics are predictive. Homophily and baseline task learnability anticipate which direction of transfer is worth attempting; edge homophily ranges from 0.05 (roman) to 0.81 (Cora) in the reported table, with Usa at 0.70 and Europe at 0.45.

Methodology in Plain English

The authors start by fixing an evaluation interface that both tasks share. All nodes and features are available, but only training-node labels are used to train NC, with validation and test labels held out. Edges are split into LP positives, and a single observed adjacency is built from the training positives only — every embedding used for NC training, NC evaluation, and LP evaluation must come from message passing on that same observed graph, so validation and test edges never appear in the neighborhoods. LP negatives are generated once per dataset with fixed seeds at a 1:1 positive-to-negative ratio and reused across all methods and runs, so that differences between methods reflect transfer rather than negative sampling. NC label splits use 60%/20%/20% and LP positive edge splits use 80%/10%/10%, held constant everywhere.

Five transfer regimes span a controlled spectrum. Warm Start (WS) initializes the target encoder with source-task encoder weights and trains a fresh, randomly initialized task head. Embedding Transfer (ET) freezes the source encoder, computes source embeddings, and either replaces the node features with them (ET-Rep) or concatenates them with the original features (ET-Concat), then trains a full target encoder from scratch — the transfer signal enters only through the inputs. Multi-View (MV) trains two task-specific encoders jointly with their own losses plus a symmetric InfoNCE cross-view alignment term using in-batch negatives, where only the alignment weight is tuned. Joint training uses one shared encoder with two heads and a weighted sum of the two task losses, with the task weight selected on validation.

Training uses a two-layer GNN encoder with fixed hidden dimension and dropout, learning rate fixed at 0.01, up to 200 epochs, and early stopping on validation. Checkpoint selection for Joint uses a combined validation criterion based on relative improvements over single-task baselines. For reporting, the CoTask Score normalizes each task against a fixed reference model — a standard GCN trained under the protocol — taking the average of the percentage gain in NC accuracy and the percentage gain in LP AUC, so the two tasks can be summarized on one unitless scale. The same construction is noted to apply if accuracy and AUC are replaced by macro-F1 for NC or Hits@K for LP.

Experiments span 11 datasets: homophilic citation graphs (Cora, Citeseer, PubMed), heterophilic graphs (Texas, Wisconsin, Cornell, Actor, roman-empire), and mixed structure-dominant graphs (Usa, Europe, Brazil). Main-text results are transductive, with an inductive variant described in an appendix.

Why This Matters

• Impact on research. Prior graph benchmarks standardize evaluation within a single task category but not across coexisting supervision signals. This work supplies a common, leakage-free interface for NC, LP, and the transfer between them, and shows that reported cross-task gains can be protocol artifacts unless splits, the message-passing graph, and LP negatives are fixed.
• Real-world applications:
  • Social networks. Classifying whether a node is a human user or a bot with node labels, while predicting potential friendship connections with edge supervision — the paper's own example of two tasks living on one graph.
  • Entity-centric systems. Platforms that classify entities, predict missing or future links, rank candidates, and detect anomalies over the same underlying structure.
  • Low-homophily or geographic/network datasets. Settings such as Usa, Europe, and Brazil, where edge structure is highly informative but labels are not aligned with connectivity, are exactly where LP-to-NC reuse as structural pretraining can pay off even when homophily is low.
  • Cost-sensitive deployments. Systems that must amortize training and serving costs by keeping a single shared encoder and a single embedding space rather than optimizing each objective separately.
• Industry relevance. The CoTask Score gives practitioners one number for joint NC+LP utility when one model must serve both tasks, and the homophily-based predictors offer a cheap prior on whether a transfer attempt is likely to help — or to cause the negative transfer that ET-Rep routinely produced on homophilic graphs.

Future Directions

1. Extend to other backbone families. The paper explicitly limits its claims to GCN, GraphSAGE, and GPS, and flags newer sequence-inspired or state-space graph architectures as an important frontier where the patterns remain untested.
2. Broaden to other task combinations. Graph classification, community detection, and subgraph-level prediction raise harder protocol questions about leakage, overlapping supervision structures, and compatible splits, which the authors position as non-trivial extensions requiring separate treatment.
3. Connect to cross-graph and cross-domain transfer. Same-graph transfer is presented as orthogonal to generalization from source graphs to unseen target graphs; understanding the same-graph case under controlled conditions is framed as a prerequisite before asking whether these patterns persist under distribution shift.
4. Make transfer guidance more operational. The link from homophily and baseline task learnability to a recommended transfer mechanism (the predictor framing behind the CoTask Score) is the natural place to build a practical decision rule, including the choice of metric substitutions such as macro-F1 or Hits@K.

Target Audience

Researchers and practitioners working on graph neural networks who need to evaluate or deploy models on graphs where node labels and edges are both available. It is most useful to those designing multi-task or shared-encoder graph systems, to benchmark authors who want leakage-free cross-task protocols, and to applied machine learning engineers deciding whether reusing supervision across node classification and link prediction is worth the effort — or likely to cause negative transfer. Readers without background in GNN evaluation practice and link prediction sampling will need to consult cited prior work first.

Note: the provided paper content is truncated during Section 5.1, so the paper's stated limitations section, the full inductive results (Appendix A.4), and the appendix-level implementation details referenced in the text are not available in the supplied material and are not summarized above.

Authors’ abstract

Many real-world graphs support multiple predictive tasks over the same underlying structure, creating an opportunity to reuse supervision across node classification (NC) and link prediction (LP). However, existing evaluations often rely on incompatible splits, observed-graph assumptions, and negative sampling rules, making conclusions about same-graph cross-task transfer unreliable. We formalize same-graph NC-LP transfer and propose a leakage-free protocol that fixes node and edge splits, uses a shared message-passing graph that excludes evaluated edges, and employs fixed negatives for LP. Across three backbones (GCN, GraphSAGE, GPS), we find that transfer is strongly directional and predictable: NC $\to$ LP is consistently beneficial on homophilic graphs, while LP $\to$ NC is fragile and can even degrade accuracy under naive representation reuse. LP $\to$ NC becomes reliably positive mainly in a structure-dominant regime where LP is easy but NC is unsaturated, suggesting that LP acts as structural pretraining. Finally, we introduce the CoTask Score (CTS) to summarize joint NC+LP utility when a shared encoder must serve both tasks, and show that simple dataset statistics, especially homophily, can guide mechanism choice and help avoid negative transfer.