🤖 machine learning

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

This paper establishes a leakage-free evaluation protocol for same-graph cross-task transfer between node classification and link prediction in GNNs, revealing that transfer is strongly directional and predictable based on graph homophily, and introduces the CoTask Score to guide mechanism selection and avoid negative transfer.

Original authors: Neelam Akula, Surbhi Kumar, Murat Kantarcioglu, Baris Coskunuzer

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Neelam Akula, Surbhi Kumar, Murat Kantarcioglu, Baris Coskunuzer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve mysteries in a giant, bustling city. In this city, the "graph" is the map of all the streets and connections between people. Usually, detectives specialize in just one type of case: either figuring out who a person is (like, "Is this person a baker or a baker's assistant?") or predicting who will become friends with whom next ("Will Alice and Bob hang out?"). For a long time, scientists building computer brains to solve these puzzles treated these as two totally separate jobs, using different rulebooks for each. But in the real world, the city map is the same for both jobs, and the clues for one often help with the other. The big question is: if you teach your computer brain to solve the "who is who" puzzle first, will it automatically get better at the "who will be friends" puzzle? Or does the reverse work? And more importantly, could trying to mix these clues actually confuse the computer and make it worse at both?

This paper dives into that exact question, but with a very strict set of rules to make sure the answer is real and not just a trick of the experiment. The researchers set up a "leakage-free" protocol, which is like making sure the detective doesn't accidentally peek at the answer key while studying the clues. They tested this on three different types of computer brains (called GCN, GraphSAGE, and GPS) across eleven different city maps. What they found is that the relationship between these two tasks is not a fair fight; it's a one-way street that depends entirely on how the city is built.

Here is the surprising twist they uncovered: Teaching the computer to identify people first (Node Classification) almost always helps it predict friendships later (Link Prediction), but only if the city is "homophilic." Think of homophily as a neighborhood where people who look alike or have the same job tend to live on the same street. In these friendly neighborhoods, knowing who lives where gives you a huge hint about who will become friends. It's like knowing that all the bakers live on Baker Street; if you see two people on that street, you can guess they might be friends.

However, the reverse is much trickier. If you teach the computer to predict friendships first and then try to use that knowledge to identify people, it often backfires. In fact, in many cases, it makes the computer worse at identifying people than if it had just started from scratch. The authors found that this "friendship-first" approach only works in a very specific, rare type of city: one where the street layout is so obvious that predicting friendships is easy, but figuring out who people are is still a mystery. In these "structure-dominant" cities, the friendship clues act like a structural map that helps the computer learn the layout, which then helps it guess the people's identities. But if the city is messy or the clues are weak, trying to reuse the friendship knowledge is like trying to drive a car using a map of a different country—it just leads to a crash.

The researchers also introduced a new way to measure success called the "CoTask Score." Imagine you are running a bakery and a coffee shop in the same building. You want to know if hiring one person to do both jobs saves money without ruining the quality of either. The CoTask Score is their way of calculating if the combined effort is actually better than doing the two jobs separately. They found that the safest bet is usually to train the computer to do both jobs at the same time, rather than trying to reuse a finished product from one job for the other.

In short, the paper suggests that while reusing knowledge between these two tasks is possible, it's not a magic bullet. You can't just assume that solving one problem will automatically solve the other. The direction matters, and the type of data matters. If your data is full of similar neighbors, teaching the computer to identify them first is a great shortcut. But if you try to teach it to predict friendships first to help with identification, you might just be wasting your time or making things worse. The key takeaway is that before you try to mix these tasks, you need to check your "city map" to see which direction the wind is blowing, or you might end up flying the wrong way.

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

1. Problem Statement

Real-world graph applications often involve multiple predictive tasks operating on the same underlying structure, specifically Node Classification (NC) and Link Prediction (LP). While these tasks frequently coexist on a single dataset (e.g., social networks requiring both user classification and friend recommendation), existing research often evaluates them in isolation.

The paper identifies a critical gap: same-graph cross-task transfer is poorly understood because standard evaluation protocols are incompatible. Existing pipelines often use:

Different data splits for NC and LP.
Different observed graphs for message passing (potentially leaking evaluation edges into training).
Different negative sampling rules for LP.

These inconsistencies create "protocol artifacts" that can mask true transfer effects or generate false positives. The authors aim to formalize the problem of reusing supervision signals between NC and LP on a single graph under a leakage-free, standardized evaluation protocol to determine when and how transfer is beneficial.

2. Methodology

2.1 Leakage-Free Evaluation Protocol

The authors propose a rigorous protocol to ensure fair comparison and prevent information leakage:

Fixed Splits: Node splits (for NC) and edge splits (for LP positives) are fixed across all methods and runs.
Shared Observed Graph: A single adjacency matrix $A_{obs}$ is constructed using only the training edges ( $E^+_{tr}$ ). Crucially, edges used for validation or testing ( $E^+_{va}, E^+_{te}$ ) are excluded from the message-passing graph. This prevents evaluation edges from leaking into node embeddings.
Fixed Negatives: Negative edge sets for LP are generated once with fixed seeds and reused across all methods, ensuring that performance differences reflect transfer efficacy rather than sampling variations.
Metrics: The paper introduces the CoTask Score (CTS), a normalized metric that averages the percentage gain of a model over a fixed reference (a standard GCN) across both NC accuracy and LP AUC. This allows for a unified assessment of multi-objective performance.

2.2 Transfer Strategies

The study evaluates five lightweight cross-task transfer strategies across three backbone architectures (GCN, GraphSAGE, GPS):

Warm Start (WS): Initializes the target task's encoder with weights learned from the source task, while re-initializing the task-specific head.
Embedding Transfer - Replace (ET-Rep): Freezes source embeddings and uses them as the sole input features for the target task encoder.
Embedding Transfer - Concat (ET-Concat): Concatenates original node features with frozen source embeddings as input to the target encoder.
Multi-View (MV): Trains two separate task-specific encoders simultaneously with an explicit cross-view alignment loss (InfoNCE) to encourage compatible representations.
Joint Training (Joint): Uses a single shared encoder optimized simultaneously for both NC and LP losses with task-specific heads.

2.3 Experimental Setup

Experiments were conducted on 11 datasets spanning three regimes:

Homophilic: Citation graphs (CORA, CITESEER, PUBMED) where edges connect similar labels.
Heterophilic: Graphs where edges connect dissimilar labels (TEXAS, WISCONSIN, ACTOR, etc.).
Structure-Dominant: Graphs where LP is highly learnable from structure, but NC has significant "headroom" (USA, EUROPE, BRAZIL).

3. Key Findings

3.1 Directional Asymmetry

The most significant finding is that cross-task transfer is strongly directional and asymmetric:

NC $\to$ LP: This direction is consistently beneficial on homophilic graphs. When nodes with the same label are connected, NC supervision shapes embeddings that inherently encode the inductive bias required for edge scoring.
LP $\to$ NC: This direction is fragile. In homophilic graphs, naive reuse of LP representations often degrades NC accuracy (negative transfer). However, LP $\to$ NC becomes reliably positive in structure-dominant regimes (e.g., USA, EUROPE, BRAZIL). In these cases, LP acts as structural pretraining, improving NC accuracy even when homophily is low, provided the NC task is unsaturated.

3.2 Predictors of Success

The paper demonstrates that simple dataset statistics can predict transfer outcomes:

Homophily ( $H_e, H_n$ ): Strongly predicts gains for NC $\to$ LP. Higher homophily correlates with better transfer performance.
Structure-Dominance: Predicts gains for LP $\to$ NC. This regime is characterized by high baseline LP learnability and low baseline NC accuracy.
Mechanism Choice:
- Joint and Multi-View (MV) training are generally the most stable and robust strategies, avoiding the high variance seen in embedding transfer methods.
- Embedding Transfer (ET-Rep) exhibits the highest variance, often leading to severe negative transfer, particularly in the LP $\to$ NC direction.

3.3 CoTask Score (CTS) Analysis

The CTS analysis confirms that while transfer can improve one task, it may harm the other. The "best" strategy depends on the regime:

On homophilic graphs, coupled training (MV, Joint) yields the highest combined utility.
On heterophilic graphs, coupled training remains the safest choice, whereas post-hoc reuse (WS, ET) is unreliable.
In structure-dominant regimes, specific mechanisms (like ET-Concat or Joint) can yield net positive gains, but this is not automatic.

4. Contributions

Problem Formalization: Defined same-graph cross-task learning between NC and LP with coexisting supervision signals.
Leakage-Free Protocol: Introduced a standardized evaluation interface that fixes splits, observed graphs, and negative sampling to enable reproducible, fair comparisons.
Systematic Study: Evaluated five transfer strategies across three backbones, documenting a robust directional asymmetry that persists across architectures.
Predictors and Metrics: Introduced the CoTask Score (CTS) for multi-objective evaluation and identified homophily and structure-dominance as interpretable diagnostics for predicting transfer success.

5. Significance and Claims

The paper claims to provide a clean evaluation paradigm for same-graph transfer learning, moving beyond the "protocol artifacts" that have plagued previous studies. Its primary contribution is not a new architecture, but rather actionable guidance for practitioners:

Transfer should not be assumed to be symmetric; NC $\to$ LP and LP $\to$ NC have distinct drivers.
Homophily is a reliable predictor for NC $\to$ LP gains.
Structure-dominance (high LP learnability, low NC saturation) is the key regime for LP $\to$ NC gains.
Coupled training (Joint/MV) is generally safer and more stable than naive representation reuse (Warm Start/Embedding Transfer), especially in heterophilic or mixed regimes.

The authors emphasize that these findings allow practitioners to decide when and which direction of reuse is worth attempting, and when a coupled regime offers a superior single-model solution for both tasks, thereby reducing training and serving costs without relying on misleading gains from protocol inconsistencies.