A graph neural network (GNN) is a model that predicts properties of nodes or whole graphs by having each node repeatedly collect information from its neighbors. A common way to build one for a new task, such as molecular toxicity, is to take a GNN pre-trained on a large molecule corpus and fine-tune it with low-rank adaptation (LoRA). LoRA freezes the original weights and learns a small additive correction built from a few rank-one pieces. Current explanation methods look at the adapted model as one opaque network. They report which atoms and bonds mattered, but they cannot say whether that evidence came from what the model already knew after pre-training or from what the fine-tuning update added. That question matters in practice. A shortcut the model picked up during fine-tuning looks just like a genuine toxicophore in an ordinary saliency map.
We propose to use LoRA's additive structure to answer the question directly. Every adapted layer computes $h = W_0x + BAx$, where $W_0$ holds the frozen pre-trained weights and $BA$ is the learned low-rank update. Relevance propagation carries a prediction score backwards through the network while keeping its total fixed. We can therefore split relevance at each layer between the frozen path and the adapter path, and further among the adapter's $r$ rank-one components. Carried through to the input, the result is an exact decomposition of an explanation into a pre-training part and a set of subgraphs that fine-tuning made relevant, one per rank component. We call these "delta explanations," and we test them against ground truth that we construct so that the answer to "what was added" is known.
Core Claim
For a GNN fine-tuned with a low-rank adapter, a conservation-preserving attribution can be split exactly by computational pathway, with no extra model training. The adapter-routed share of relevance recovers the input structures that the fine-tuning task introduced. It does so more accurately than the obvious alternative, which is to subtract an explanation of the pre-trained model from an explanation of the fine-tuned model. Each rank-one component of the adapter maps to a coherent subgraph pattern whose removal changes the prediction by roughly the amount its relevance share predicts. A secondary claim follows from the same measurement. The finding in Evaluating Attribution for Graph Neural Networks that attribution accuracy decays late in training while predictive accuracy holds steady can be localized. Under LoRA fine-tuning, we predict the decay concentrates in the adapter pathway, which gives a concrete signal for early stopping.
Method
Two papers supply the method. Efficient Computation of Higher-Order Subgraph Attribution via Message Passing (sGNN-LRP) shows that layer-wise relevance over walks, meaning paths of nodes traced through the layers, is equivalent to marginal probabilities in a Markov chain. It computes them by message passing in time linear in depth rather than exponential. We extend the chain's state from "node at layer $l$" to "node at layer $l$ with pathway $p$," where $p$ is either the frozen path or one of the adapter's $r$ rank components. At a layer, the contribution of input unit $i$ to output unit $j$ decomposes as
$$z_{ij} = W_{0,ji}\,x_i + \sum_{k=1}^{r} B_{jk}A_{ki}\,x_i ,$$
and the standard stabilized LRP rule distributes relevance in proportion to each term. Because the rule conserves total relevance, the pathway shares sum exactly to the original explanation. Tracking the pathway multiplies the cost of message passing by roughly $r+1$, so it stays cheap for the ranks LoRA uses.
The second source is Graph Low-Rank Adapters of High Regularity for Graph Neural Networks and Graph Transformers, which introduced GConv-Adapter. That adapter wraps its low-rank bottleneck in normalized graph convolutions, so its path carries messages up to two extra hops. In our construction those hops become extra walk steps that exist only in the adapter pathway. This lets us ask a question that plain LoRA cannot pose: did fine-tuning make the model use longer-range structure that the backbone ignored?
We apply the method to backbones and adapters exactly as that paper trains them: a GIN pre-trained on ZINC15, with LoRA or GConv-Adapter fine-tuning on the molecular benchmarks. That paper reports only predictive accuracy and stability. Our work is the interpretability layer it leaves open.
Research Program
The program has three parts.
The first part builds benchmarks where the "delta" has a known answer. Existing ground-truth benchmarks only say what matters for a label. We build staged versions in which the pre-training task and the fine-tuning task depend on partly different motifs, so the structures that fine-tuning should add are known.
On synthetic graphs, we pre-train on BA-2motifs-style classification where the house motif decides the label. We then fine-tune on a task where the label depends on house AND cycle, so the cycle is the ground-truth addition. A node-level analogue uses BA-Shapes, Tree-Cycles, and Tree-Grid.
On molecules, the chemical-group tasks from Evaluating Attribution for Graph Neural Networks provide computable ground truth. We pre-train a head-free backbone on ZINC15, fine-tune first on single-group detection such as Benzene, and then fine-tune an adapter on the conjunctive Amine AND Ether task. This setup mirrors the synthetic one.
The second part turns the per-component decomposition into readable explanations. For each rank component $k$, we aggregate its input relevance across a dataset and summarize it as motifs, using the candidate-motif extraction of MotifExplainer as a vocabulary. The goal is statements of the form "component 3 of the Tox21 adapter fires on nitro groups attached to aromatic rings."
The third part tracks how pathway shares and ground-truth recovery evolve across fine-tuning epochs. This tests the localization claim about late-training decay.
Baselines
The naive reference points come first. Probing GNN Explainers shows that a random-edge explanation often achieves lower unfaithfulness than published explainers. We include it, applied only to the edges that differ in relevance between the pre-trained and fine-tuned model, as a floor every method must clear. We also include "no split": the whole-model sGNN-LRP explanation of the fine-tuned model, scored against the delta ground truth. This measures how much a method gains by separating pathways at all.
The strongest competitor for recovering what was added is difference of explanations. We attach the fine-tuned prediction head to the frozen backbone, attribute both models with the same method, and subtract. We run it with three attribution methods: sGNN-LRP; CAM, which Evaluating Attribution for Graph Neural Networks found the most accurate, stable, and faithful attribution on GCN; and Integrated Gradients. It sees the same models, the same graphs, and a matched hyperparameter search over explanation sparsity. We also run PGExplainer (Parameterized Explainer for Graph Neural Network), trained separately on each model and then subtracted, as a learned-explainer variant. It gets the same training-graph budget and its recommended tuning.
For the strongest numbers, the file gives only indirect figures. PGExplainer reports a 24.7% relative AUC improvement over GNNExplainer on BA-2motifs. For CAM, no reported numerical AUROC was found on Lacuna, only its top ranking. For the delta task itself, no reported result was found on Lacuna, because the task is new. Beating the competitor means higher AUROC against the added-motif ground truth, together with higher deletion faithfulness, with the gap exceeding seed-to-seed spread over at least five fine-tuning seeds.
On the predictive side, adapters must remain competitive. GConv-Adapter reports an 81.142% average ROC-AUC across the molecular classification tasks and a 9.49% average RMSE reduction relative to full fine-tuning. We reproduce these numbers before interpreting any explanation. Vanilla LoRA gets the learning-rate sweep that A Unified Study of LoRA Variants shows is necessary.
Two ablations isolate the key ingredients. The first applies the same pathway split to full fine-tuning, writing $W = W_0 + \Delta W$ with a full-rank $\Delta W$. It tests whether the additive split alone does the work. The second compares truncated SVD components of that full $\Delta W$ against LoRA's native rank components. It tests whether low rank specifically produces cleaner per-component concepts.
Why not just…?
Why not just subtract the pre-trained model's explanation from the fine-tuned one? This is the real threat. The subtraction breaks down in a specific way. The pre-trained backbone with the new head was never trained to make the fine-tuning prediction, so its explanation describes a different function. And whenever the added motif matters only together with the old one, as in the conjunctive Amine AND Ether and house-AND-cycle tasks, subtracting two nonlinear attributions spreads the interaction across both motifs. Pathway splitting avoids both problems because it attributes a single forward pass of the model actually being deployed. If subtraction matches us on the conjunctive staged benchmarks, the method adds little, and we will report that plainly.
Why not just use attention weights or a GAT to see what changed? Evaluating Attribution for Graph Neural Networks found attention weights to be poor indicators of importance, well behind CAM and Integrated Gradients. The GIN backbones used for pre-training have no attention in any case. We include attention-difference on a GAT variant only as a sanity row.
Why not just treat the adapted model as a black box and fit a local sparse surrogate? Explaining Graph Neural Networks Using Interpretable Local Surrogates (ILS) is model-agnostic and strong on synthetic benchmarks, reaching 96.61% on BA-GraphR regression. But a surrogate sees only inputs and outputs, so it can recover the delta only through the same subtraction, with the same weaknesses. It also cannot attribute to adapter components at all. We run ILS-difference in the staged benchmarks to show this empirically.
Why not just use full fine-tuning and decompose the weight difference? Our framing already admits that this might work. The additive split applies to any $W_0 + \Delta W$. The claim specific to LoRA is that native low-rank components give sparser, more coherent per-component motifs than SVD components of a full-rank update. The second ablation settles this. If SVD matches LoRA, the contribution becomes the pathway decomposition itself, and it applies to any fine-tuned GNN.
Proposed Experiments and Decisive Checks
The first check is correctness. On every model, the pathway shares must sum to the whole-model sGNN-LRP explanation up to numerical tolerance. The split must also remain cheap: we measure runtime against the reported 29.94 ms per sGNN-LRP explanation on Graph-SST2, expecting roughly an $(r+1)$-fold slowdown.
The second check, and the decisive one, is recovery of the delta. On the staged synthetic and chemical-group benchmarks, adapter-pathway relevance should score higher AUROC against the added-motif ground truth than the difference-of-explanations baselines, with the largest margin on conjunctive tasks.
The third check is causal. For each rank component $k$, we remove it by zeroing column $k$ of $B$. We then compare the resulting logit change with the relevance share assigned to component $k$ across the test set. We also remove the top adapter-attributed edges and measure the drop in the adapter's contribution to the logit, in the spirit of the PARC faithfulness score from Evaluating Attribution for Graph Neural Networks. A high rank correlation between predicted and observed effects is the result that would make per-component explanations trustworthy. A weak correlation would falsify the component-level claim even if pathway-level recovery holds.
The fourth check covers real molecular tasks with known chemistry: MUTAG, where nitro groups and carbon rings are the reference toxicophores, and Tox21, ClinTox, BACE, and HIV. We test whether adapter-attributed motifs correspond to task-relevant groups that the backbone alone does not emphasize. We also run a planted-shortcut variant, adding a spurious marker atom correlated with the label in fine-tuning data only, to see whether the adapter pathway isolates the shortcut. We compare plain LoRA against GConv-Adapter, measuring how much adapter relevance travels over the adapter's extra hops.
The last check tracks recovery accuracy per pathway across epochs to test whether late-training attribution decay lives in the adapter path. We also measure stability under small graph perturbations, following Probing GNN Explainers. This asks whether the bounded Lipschitz constant proven for GConv-Adapter goes with more stable adapter-pathway explanations than those from plain LoRA.
References
- Evaluating Attribution for Graph Neural Networks
- Efficient Computation of Higher-Order Subgraph Attribution via Message Passing
- Graph Low-Rank Adapters of High Regularity for Graph Neural Networks and Graph Transformers
- MotifExplainer: a Motif-based Graph Neural Network Explainer
- Probing GNN Explainers: A Rigorous Theoretical and Empirical Analysis of GNN Explanation Methods
- Parameterized Explainer for Graph Neural Network
- A Unified Study of LoRA Variants: Taxonomy, Review, Codebase, and Empirical Evaluation
- Explaining Graph Neural Networks Using Interpretable Local Surrogates