Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance

# Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance
Source: [https://arxiv.org/html/2608.14121](https://arxiv.org/html/2608.14121)
Taraneh Younesian \(✉\)Steve AzzolinAffiliation:University of Trento, ItalyE\-mail[steve\.azzolin,francesco\.ferrini,vincenzomarco\.deluca,stefano\.teso@unitn\.it](mailto:steve.azzolin,francesco.ferrini,vincenzomarco.deluca,[email protected])Antonio LongaAffiliation:UiT The Arctic University of Norway, NorwayE\-mail[\[antonio\.longa@uit\.no\]\(mailto:antonio\.longa@uit\.no\)](mailto:[[email protected]](mailto:[email protected]))Francesco FerriniAffiliation:University of Trento, ItalyE\-mail[steve\.azzolin,francesco\.ferrini,vincenzomarco\.deluca,stefano\.teso@unitn\.it](mailto:steve.azzolin,francesco.ferrini,vincenzomarco.deluca,[email protected])Vincenzo Marco De LucaAffiliation:University of Trento, ItalyE\-mail[steve\.azzolin,francesco\.ferrini,vincenzomarco\.deluca,stefano\.teso@unitn\.it](mailto:steve.azzolin,francesco.ferrini,vincenzomarco.deluca,[email protected])Stefano TesoAffiliation:University of Trento, ItalyE\-mail[steve\.azzolin,francesco\.ferrini,vincenzomarco\.deluca,stefano\.teso@unitn\.it](mailto:steve.azzolin,francesco.ferrini,vincenzomarco.deluca,[email protected])

###### Abstract

Graph Neural Networks \(GNNs\) can solve prediction tasks by unintentionally exploitingshortcuts—that is, edges, nodes, and features that correlate with but are not causal for the prediction—which compromise their reliability in out\-of\-distribution tasks\. We introduceXIGL\(eXplanatory Interactive Graph shortcut unLearning\), an architecture\-agnostic human\-in\-the\-loop strategy for removing such shortcuts from GNNs\. Our key insight is twofold\. On the one hand, reliance on shortcuts can be detected by inspecting GNN explanations\. On the other hand, once made aware of such shortcuts, sufficiently expert users can provide tailored corrective feedback, which helps deconfound the model\.XIGLsupports any query strategy; however, since corrective feedback can be expensive to acquire, we develop an active learning strategy for prioritizing explanations that are more likely to display shortcut behavior, lowering annotation and cognitive costs\. We showcase the effectiveness ofXIGL, including both existing and proposed explanation\-based strategies, on several GNN architectures\. Our implementation is available online\.111https://github\.com/TYounesian/xilgraph\.git

###### Keywords:

Graph Neural Networks Explainable AI Active Learning Shortcut Learning Deconfounding\.

## 1Introduction

Graph Neural Network \(GNN\) classifiers can pick up onshortcuts, also calledconfounders, patterns that happen to correlate with the label in\-distribution, allowing the model to achieve high accuracy but not generalize outside of it\. Reliance on such shortcuts—including watermarks\[[11](https://arxiv.org/html/2608.14121#bib.bib28)\], metadata\[[6](https://arxiv.org/html/2608.14121#bib.bib29)\], and simple input statistics\[[16](https://arxiv.org/html/2608.14121#bib.bib16)\]—can prevent models from behaving properly upon deployment when the data distribution changes\. Since providing more \(observational\) training data is insufficient to resolve shortcuts, existing works address them by employing other kinds of data, such as examples from multiple domains\[[1](https://arxiv.org/html/2608.14121#bib.bib14),[10](https://arxiv.org/html/2608.14121#bib.bib15)\], that are not always available\.

We take a different route\. Specifically, we exploit the fact that, by unveiling the GNN’s reasoning process, GNNexplanationscan naturally expose its reliance on shortcuts\[[2](https://arxiv.org/html/2608.14121#bib.bib17),[6](https://arxiv.org/html/2608.14121#bib.bib29),[11](https://arxiv.org/html/2608.14121#bib.bib28),[16](https://arxiv.org/html/2608.14121#bib.bib16),[18](https://arxiv.org/html/2608.14121#bib.bib26),[5](https://arxiv.org/html/2608.14121#bib.bib2)\]\. We then designXIGL\(eXplanatory Interactive Graph shortcut unLearning\), an active learning pipeline, illustrated in Figure[1](https://arxiv.org/html/2608.14121#S1.F1), that iterativelyi\) employs an intuitive selection strategy to identify those instances whose explanations are more likely to reveal shortcut behavior;ii\) requests a human annotator to correct these explanations by indicating what nodes of the input graph the GNN shouldnotrely on; andiii\) adapts the GNN based on the user’s corrections through an end\-to\-end differentiable loss function geared toward shortcut removal\[[18](https://arxiv.org/html/2608.14121#bib.bib26),[23](https://arxiv.org/html/2608.14121#bib.bib27)\]\.

Compared to existing work on explanation\-based deconfounding of GNNs\[[28](https://arxiv.org/html/2608.14121#bib.bib5)\], which assumes that explanatory supervision is available for the entire training set,XIGLactively prioritizes annotating a subset of the data with those explanations that better capture dependency on shortcuts\. This avoids the need to annotate examples where the model is behaving properly, while focusing the annotator’s effort on subgraphs for which feedback is needed and useful, thus reducing annotation and cognitive costs\.

Focusing on graph classification tasks, whileXIGLsupports any query strategy, our experiments demonstrate howXIGL’s explanation\-based query strategies mostly improve on existing active learning baselines, which select instances based purely on prediction\-based informativeness across several GNN architectures\.

Figure 1:Overview ofXIGL\. We first extract explanations from the GNN being trained over a small subset of the dataset; an active learning strategy then selects the most informative explanations for an expert to correct by indicating which nodes the model should not rely on and which nodes it should rely on instead; the corrections then fine\-tune the GNN via an explanation\-guided loss that removes shortcut dependence\. The process iterates until the annotation budget is reached\.
## 2Preliminaries

LetG=\(𝒱,ℰ,X\)G=\(\\mathcal\{V\},\\mathcal\{E\},X\)be an \(undirected\)graph, where𝒱=\{v1,…,vN\}\\mathcal\{V\}=\\\{v\_\{1\},\.\.\.,v\_\{N\}\\\}is the set of nodes,ℰ⊆𝒱×𝒱\\mathcal\{E\}\\subseteq\\mathcal\{V\}\\times\\mathcal\{V\}is the set of edges,X∈ℝN×dX\\in\\mathbb\{R\}^\{N\\times d\}is the node features, andy∈\{1,…,l\}y\\in\\\{1,\\ldots,l\\\}is the label assigned toGG\. We tacklegraph classificationproblems where the goal is to learn a functionf:𝒢→𝒴f:\\mathcal\{G\\to\\mathcal\{Y\}\}mapping from the set of possible graphs𝒢\\mathcal\{G\}to the set of labels𝒴\\mathcal\{Y\}from training examples of annotated graphs\.

We focus on scenarios in which the input graphsGGencode both “causal” and “spurious” subgraphs, denotedCCandSS, respectively: whereasCCdetermines the ground\-truth labelyy,SSonly correlates with it\. To build intuition, considerCPatchMNIST\[[4](https://arxiv.org/html/2608.14121#bib.bib1)\]: here, the graphsGGrepresent handwritten digits in terms of superpixels \(nodes\) and their adjacency relation \(edges\)\. Node features encode the average color of the corresponding superpixel, and the labelyyis the digit itself\. Crucially, the graphs are manipulated such that the color of the rightmost superpixels correlates with the digit, and as such act as shortcuts\.

Neural nets are known to be biased to learn simpler solutions\[[26](https://arxiv.org/html/2608.14121#bib.bib4)\]\. In real\-world settings, the confounderSSis often easier to learn than the causal oneSS\. In our example, since the spurious superpixels are highly discriminative, GNNs tend to exploit them for prediction\. This yields excellent in\-distribution performance but near\-random accuracy in unseen test instances where color is removed\.

Throughout, we refer to the nodes in the causal subgraphCCasrelevantnodes and denote them𝒱r\\mathcal\{V\}\_\{r\}, and to the nodes of its complement asirrelevantnodes, denoting them𝒱i​r\\mathcal\{V\}\_\{ir\}\. The latter covers the spurious subgraphSS, as well as the rest of the graph\. Ideally, we wish our GNNs to make predictions for the “right reason”, that is, by relying on the causal subgraph rather than the spurious one\.

## 3Deconfounding through Active Explanation

We aim to deconfound the model by guiding it to not use shortcuts\. We do so by penalizing the impact of irrelevant subgraphs of nodesVi​rV\_\{ir\}on the model\. As an indication of the impact of nodes, we use input gradients\[[17](https://arxiv.org/html/2608.14121#bib.bib18),[28](https://arxiv.org/html/2608.14121#bib.bib5)\], which we adopt among the many possible GNN explainers \(see\[[14](https://arxiv.org/html/2608.14121#bib.bib20),[27](https://arxiv.org/html/2608.14121#bib.bib19)\]for a survey\) for their simplicity and because they do not require training an ad\-hoc explainer\. We then minimize this value for irrelevant nodes in each graph, as follows:

ℒ\\displaystyle\\mathcal\{L\}=ℒc​e\+λℒe=−∑i=1ly^ilog\(y^i\)\+λ∑j=1N\(1−ej\)​sj2∑j=1Nsj2,\\displaystyle=\\mathcal\{L\}\_\{ce\}\+\\lambda\\mathcal\{L\}\_\{e\}=\-\\sum\_\{i=1\}^\{l\}\\hat\{y\}\_\{i\}\\log\(\\hat\{y\}\_\{i\}\)\+\\lambda\\frac\{\\sum\_\{j=1\}^\{N\}\(1\-e\_\{j\}\)s\_\{j\}^\{2\}\}\{\\sum\_\{j=1\}^\{N\}s\_\{j\}^\{2\}\},\(1\)wherey^i\\hat\{y\}\_\{i\}is the predicted probability of classii,sj=∑i=1l∇j​log​\(y^i\)s\_\{j\}=\\sum\_\{i=1\}^\{l\}\\nabla\_\{j\}\\log\(\\hat\{y\}\_\{i\}\)is sum of the input gradients,λ\\lambdais a scaling term to balance between classification and explanation loss, and𝐞∈\{0,1\}N\\mathbf\{e\}\\in\\\{0,1\\\}^\{N\}is the ground\-truth explanation of the graphGG\. In particular, for each nodejj,𝐞\\mathbf\{e\}indicates ifjjbelongs to the relevant or irrelevant subgraph, i\.e\.,eje\_\{j\}is11ifvjv\_\{j\}is relevant, i\.e\.,vj∈𝒱rv\_\{j\}\\in\\mathcal\{V\}\_\{r\}, and is00otherwise\. In words, minimizingℒe\\mathcal\{L\}\_\{e\}corresponds to minimizing the ratio of the input gradients of irrelevant nodes and the total input gradients across all nodes\. In contrast to\[[28](https://arxiv.org/html/2608.14121#bib.bib5)\], we incorporate a normalization term in the denominator to prevent the trivial solution in which the input gradients of all nodes, including the relevant ones, collapse to zero\. The loss function above aims to balance learning the labels with learning the truly relevant subgraph\. Computing the total loss gradient during backpropagation yields second\-order gradients with respect to the input\. In practice, the resulting computational overhead was small\.

### 3\.1Passive vs\. Active Explanation Supervision

So far, we assumed access to the ground\-truth explanation𝐞\\mathbf\{e\}for every graph in the training set\. We call this scenario*passive explanation supervision*, since our model passively uses all instances to learn from\. Obtaining ground\-truth explanations for every instance is costly and time\-intensive, as it requires expert annotation to determine the task\-relevant nodes in each graph\. Therefore, this assumption is impractical in real\-world scenarios\.

To address this limitation, we introduce an annotation budgetBB, which restricts the total number of graphs for which ground\-truth explanations222We use*ground\-truth explanation*for simplicity\. In practice, the explanation corrections provided by users do not need to match the true causal mechanism in the data, as they may not know the actual cause of the phenomenon being modeled\. Yet, they may reliably indicate which nodes the model*should*rely on to avoid clear shortcuts\.can be collected\. Instead of annotating the entire dataset, we select a small, informative subset of graphs so that training on this subset yields a highly accurate model that captures the underlying causal patterns\.

We employ*active learning*\(AL\)\[[19](https://arxiv.org/html/2608.14121#bib.bib6),[12](https://arxiv.org/html/2608.14121#bib.bib9),[20](https://arxiv.org/html/2608.14121#bib.bib7),[29](https://arxiv.org/html/2608.14121#bib.bib8)\]to identify informative instances\. WhileXIGLis agnostic to the choice of query strategy and can be combined with any active learning method, we introduce two novel explanation\-based query strategies that leverage explanation uncertainty to identify instances whose explanation corrections are expected to provide the greatest benefit to the model\. We include two standard query strategies inXIGLfor comparison\.

Let𝒟U=\{\(Gi,yi\)\}i=1N\\mathcal\{D\}\_\{U\}=\\\{\(G\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{N\}be the dataset of graphs and their labels without their ground\-truth explanation\. The process begins by training the model on an initial small set ofqqgraphs for which we have the explanations denoted as𝒟exp=\{\(Gi,yi,Ei\)\}i=1q\\mathcal\{D\}\_\{\\mathrm\{exp\}\}=\\\{\(G\_\{i\},y\_\{i\},E\_\{i\}\)\\\}\_\{i=1\}^\{q\}\. Then, among the remaining rest of𝒟U\\mathcal\{D\}\_\{U\}, we select a batch of the most informative graphs and query an expert for their ground\-truth explanations\. The newly annotated batch is added to the previously annotated set𝒟exp\\mathcal\{D\}\_\{\\mathrm\{exp\}\}, and the model is subsequently fine\-tuned on this set using Eq\.[1](https://arxiv.org/html/2608.14121#S3.E1)\. We repeat this procedure forT≤BT\\leq Biterations, ensuring that the total number of annotated instances does not exceed the budget\. Algorithm[1](https://arxiv.org/html/2608.14121#alg1)shows the steps ofXIGL’s active learning\. We investigate two classes of query strategies: prediction\-based and explanation\-based\. The former are established active learning methods, whereas the latter are novel strategies proposed in this paper\. The details of each strategy are provided below:

Algorithm 1Active Learning Steps inXIGL1:Labeled graph dataset

𝒟U=\{\(Gi,yi\)\}i=1N\\mathcal\{D\}\_\{U\}=\\\{\(G\_\{i\},y\_\{i\}\)\\\}\_\{i=1\}^\{N\}, initial query budget

qq, query budget

BB, active learning iteration

TT, query strategy

𝒮\\mathcal\{S\}
2:Updated GNN model

ff
3:Select

qqgraphs and obtain their explanation correction:

𝒟exp=\{\(Gi,yi,Ei\)\}i=1q\.\\mathcal\{D\}\_\{\\mathrm\{exp\}\}=\\\{\(G\_\{i\},y\_\{i\},E\_\{i\}\)\\\}\_\{i=1\}^\{q\}\.\(2\)
4:Train an initial GNN

fθf\_\{\\theta\}using:

ℒ=ℒce\+λ​ℒe\\mathcal\{L\}=\\mathcal\{L\}\_\{\\mathrm\{ce\}\}\+\\lambda\\mathcal\{L\}\_\{\\mathrm\{e\}\}\.

5:foreach iteration

t=1,…,Tt=1,\\dots,Tdo

6:Generate predictions and explanations for graphs in

𝒟U∖𝒟exp\\mathcal\{D\}\_\{U\}\\setminus\\mathcal\{D\}\_\{\\mathrm\{exp\}\}\.

7:Select a query set

𝒬\\mathcal\{Q\}of size

B/TB/Taccording to

𝒮\\mathcal\{S\}\(Section[3\.1](https://arxiv.org/html/2608.14121#S3.SS1)\)\.

8:Obtain explanation correction for graphs in

𝒬\\mathcal\{Q\}\.

9:Update the explanation\-supervised set:

𝒟exp←𝒟exp∪\{\(Gi,yi,Ei\):Gi∈𝒬\}\.\\mathcal\{D\}\_\{\\mathrm\{exp\}\}\\leftarrow\\mathcal\{D\}\_\{\\mathrm\{exp\}\}\\cup\\\{\(G\_\{i\},y\_\{i\},E\_\{i\}\):G\_\{i\}\\in\\mathcal\{Q\}\\\}\.\(3\)
10:Fine\-tune the GNN using

ℒ=ℒce\+λ​ℒe\\mathcal\{L\}=\\mathcal\{L\}\_\{\\mathrm\{ce\}\}\+\\lambda\\mathcal\{L\}\_\{\\mathrm\{e\}\}\.

#### Prediction\-based strategies

- •Maximum Classification Entropy\(MaCE\) is one of the most popular uncertainty\-based query strategies\. MaCE queries graphs that the GNN is highly uncertain to classify, using the Shannon entropy: G∗=argmaxG∈𝒟U−∑i=1lp⁡\(yi∣G\)​log⁡p⁡\(yi∣G\)G^\{\*\}=\\argmax\_\{G\\in\\mathcal\{D\}\_\{U\}\}\-\\sum\_\{i=1\}^\{l\}p\(y\_\{i\}\\mid G\)\\log p\(y\_\{i\}\\mid G\)\(4\)
- •Random Samplingrandomly queries graphs according to a uniform distribution\. Although random sampling does not make use of predictive information, we include it in this category as a baseline for comparison\.

#### Explanation\-based strategies

- •Maximum Explanation Entropy\(MaEE\) queries the graphs that the GNN is most uncertain in itsexplanation\. Specifically, it prioritizes graphs in which input gradients are most uniform across nodes, as measured by the entropy of softmax\-normalized node\-wise input gradients: G∗=argmaxG∈𝒟U−∑i=1\|𝒱\|pi​log⁡pilog⁡\|G\|,G^\{\*\}=\\argmax\_\{G\\in\\mathcal\{D\}\_\{U\}\}\-\\frac\{\\sum\_\{i=1\}^\{\|\\mathcal\{V\}\|\}p\_\{i\}\\log p\_\{i\}\}\{\\log\|G\|\},\(5\)wherepi=exp⁡\(si\)∑j∈G​exp⁡\(sj\)p\_\{i\}=\\frac\{\\exp\(s\_\{i\}\)\}\{\\sum\{j\\in G\}\\exp\(s\_\{j\}\)\}is the softmax\-normalized input gradientsis\_\{i\}\. Since the number of nodes in each graph varies, we normalize the entropy to compare across graphs\.
- •Minimum Explanation Entropy\(MiEE\) queries the graphs that the GNN is mostcertainin its explanation\. We adopt this strategy based on the simplicity bias of neural networks\[[26](https://arxiv.org/html/2608.14121#bib.bib4)\], which causes models to preferentially learn simpler yet spurious patterns\. As a result, the model may quickly overfit to confounders and produce overly confident explanations that incorrectly attribute the prediction to the confounding features\. G∗=argminG∈𝒟U−∑i=1\|𝒱\|pi​log⁡pilog⁡\|G\|\.G^\{\*\}=\\argmin\_\{G\\in\\mathcal\{D\}\_\{U\}\}\-\\frac\{\\sum\_\{i=1\}^\{\|\\mathcal\{V\}\|\}p\_\{i\}\\log p\_\{i\}\}\{\\log\|G\|\}\.\(6\)

## 4Experiments

DatasetsWe use one synthetic and one real\-world dataset for our experiments:ER\-coloris a synthetic dataset comprising10001000Erdős\-Rényi graphs with5050nodes and edge probability of0\.050\.05\. We randomly assign the colors red, blue, green, yellow, and orange to the nodes using a uniform distribution\. We split the dataset into70%70\\%,15%15\\%, and15%15\\%for train, validation, and test sets, respectively\. We then randomly attach two motifs to the base graphs, each representing a label\. Hence, these motifs represent theright reasons\. To create confounders that induce a distribution shift between the train and validation/test sets, we addjjpurple nodes for label00andjjcyan nodes for label11as confounders only in the train set, while keeping the validation and test sets unchanged\. Node features are represented as one\-hot encoding of colors\. For confounding colors, the active feature value is set to100100rather than11, thereby amplifying the confounding signal\. We usej=1j=1for passive learning experiments, and for active learning, we randomly selectjjwherej∼Uniform​\(1,…,25\)j\\sim\\textsf\{Uniform\}\(\{1,\\ldots,25\}\)\. The reason for the varying number of confounders in active learning experiments is to create different levels of informativeness among training instances\. Figure[2](https://arxiv.org/html/2608.14121#S4.F2)shows the cases with and without confounders, i\.e\., training or validation/test sets, respectively\.

![Refer to caption](https://arxiv.org/html/2608.14121v1/images/no_confounder2.png)\(a\)No confounder,y=0y=0
![Refer to caption](https://arxiv.org/html/2608.14121v1/images/no_confounder1.png)\(b\)No confounder,y=1y=1
![Refer to caption](https://arxiv.org/html/2608.14121v1/images/confounder_nc1.png)\(c\)Confounder with a new color,y=0y=0
![Refer to caption](https://arxiv.org/html/2608.14121v1/images/confounder_nc2.png)\(d\)Confounder with a new color,y=1y=1

Figure 2:ER\-colordataset examples with and without confounders\. The trees shown are the ground\-truth explanations for each class\. As confounders, fory=0y=0, a purple node is added to the graph, and fory=1y=1, a cyan node is added\.CPatchMNIST\[[4](https://arxiv.org/html/2608.14121#bib.bib1)\]is an extension ofMNISTsp\[[9](https://arxiv.org/html/2608.14121#bib.bib3)\], itself a conversion of the MNIST dataset to graphs via a superpixelation algorithm\. InCPatchMNIST, on the other hand, the top left and bottom right superpixels are colored according to the graph label in the training set to represent confounders\. However, in validation and test sets, these colors are randomized\.

Experimental SetupWe examined four GNN architectures—GCN\[[8](https://arxiv.org/html/2608.14121#bib.bib10)\],GIN\[[25](https://arxiv.org/html/2608.14121#bib.bib12)\],GraphSAGE\[[7](https://arxiv.org/html/2608.14121#bib.bib13)\], andGAT\[[24](https://arxiv.org/html/2608.14121#bib.bib11)\]—to assess both their vulnerability to shortcut learning and their ability to recover from it by avoiding spurious dependencies through explanation guidance\. Further implementation details are available in the Appendix\. For the active learning experiments, we query55instances per round over2020rounds forER\-color, and5050instances per round over1010rounds forCPatchMNIST\. We observed that when the annotation budget is large, the gap between different query strategies largely vanishes\. Intuitively, this occurs because the informative instances selected by active learning methods constitute a subset of the data that is increasingly likely to be included in larger randomly sampled subsets\. As a result, the benefits of querying informative instances become less noticeable\.

Table 1:Effect of passive supervision inXIGLacross datasets\.Accuracy \(%\) on the unconfounded test split\. Results are averaged with standard deviation across 5 seeds\.### 4\.1Results

#### Passive Explanation Supervision

Table[1](https://arxiv.org/html/2608.14121#S4.T1)shows the test accuracy forER\-colorandCPatchMNISTfor different setups and GNN architectures, respectively\. As the results show, passive supervision generally improves the accuracy compared to no\-explanation supervision \(No Sup\.\) and their ability to generalize to unconfounded data\. We noticed that passive supervision’s performance improves further if we first train the models for a few epochs only onℒe\\mathcal\{L\}\_\{e\}, followed by training onℒc​e\+λ​ℒe\\mathcal\{L\}\_\{ce\}\+\\lambda\\mathcal\{L\}\_\{e\}, refer to Table[3](https://arxiv.org/html/2608.14121#Pt0.A1.T3)in Appendix[0\.A](https://arxiv.org/html/2608.14121#Pt0.A1)for more details\. The results shown in Table[1](https://arxiv.org/html/2608.14121#S4.T1)under Passive correspond to this setup\. As the results show, onER\-color, most models perform near the chance level without explanation supervision but improve substantially with explanation supervision\. This indicates that the models primarily rely on shortcut information and largely ignore the causal patterns, whereas explanation supervision guides them toward the causally relevant features\.

OnCPatchMNIST, most models perform slightly above chance level even without supervision, indicating that the causal signal can be learned to a limited degree\. However, explanation supervision clearly improves performance for all architectures except forGraphSAGE\. This effect is particularly evident forGIN, which quickly overfits to the shortcuts and suffers a sharp drop in test accuracy, while explanation supervision inXIGLhelps it recover and generalize better\. While early stopping can reduce overfitting, it also limits the extent to which reliance on shortcuts becomes apparent\. To expose this behavior and evaluate the robustness of explanation supervision, we do not use early stopping\.

#### Active Explanation Supervision

Table[2](https://arxiv.org/html/2608.14121#S4.T2)shows the results of the different query strategies across both datasets and all GNNs\. As the results indicate, the effectiveness of the query strategies varies across datasets and models\. Nevertheless, active query selection generally outperforms random sampling in most settings\. OnCPatchMNIST, the explanation\-based strategies consistently outperform the prediction\-based strategies across all GNNs\. Interestingly, random sampling achieves the best performance forGINin both datasets\. We leave a systematic investigation of this effect for future work\. Overall, MiEE emerges as the most consistently high\-performing query strategy across models and datasets\. These results highlight the potential of explanation\-based active learning and motivate further investigation into explanation\-driven query strategies, particularly those targeting highly confident yet potentially incorrect explanations\.

As expected, the active learning component inXIGLgenerally achieves lower performance than passive learning due to its substantially smaller annotation budget\. Nevertheless, the gap is relatively small, despite active learning using less than20%20\\%of the training data\. These results suggest that active learning can achieve performance comparable to passive learning while requiring significantly fewer annotated explanations\.

Table 2:Effect of AL strategies ofXIGLacross datasets\.Accuracy \(%\) on the unconfounded test split\. Results are averaged with standard deviation across 5 seeds\.

## 5Discussion, Related Work, & Conclusion

Our work suggests that, by exploiting expert corrections to model explanations,XIGLcan help remove shortcut dependencies from models, and that it can reduce annotation costs compared to passive baselines\. In future work, we plan to extend the experiments to more datasets and GNN architectures, and specifically to self\-explainable GNNs\[[15](https://arxiv.org/html/2608.14121#bib.bib25),[21](https://arxiv.org/html/2608.14121#bib.bib23),[22](https://arxiv.org/html/2608.14121#bib.bib24)\]\. Moreover, we plan to study the impact of explanation faithfulness\[[17](https://arxiv.org/html/2608.14121#bib.bib18)\]on debiasing success: if an explanation does not capture the model’s actual reasoning, then corrections to it may not help improve the model’s behavior\[[3](https://arxiv.org/html/2608.14121#bib.bib21)\]\. This is relevant for both post\-hoc explainers\[[13](https://arxiv.org/html/2608.14121#bib.bib22)\]and self\-explainable architectures\[[4](https://arxiv.org/html/2608.14121#bib.bib1)\]\. Additionally, evaluating different kinds of shortcuts and real\-world datasets is planned for future work\.

#### Acknowledgements

Taraneh Younesian was funded by Huawei DREAMS Lab\. All content represents the opinion of the authors, which is not necessarily shared nor endorsed by their respective employers and/or sponsors in Huawei DREAMS Lab\. Antonio Longa was supported by the Research Council of Norway through its Centre of Excellence Integreat \- The Norwegian Centre for knowledge\-driven machine learning, project number 33264\. Funded by the European Union\. Views and opinions expressed are however those of the author\(s\) only and do not necessarily reflect those of the European Union or the European Health and Digital Executive Agency \(HaDEA\)\. Neither the European Union nor the granting authority can be held responsible for them\. Grant Agreement no\. 101120763 \- TANGO

#### Disclosure of Interests\.

The authors have no competing interests to declare that are relevant to the content of this article\.

## References

- \[1\]M\. Arjovsky, L\. Bottou, I\. Gulrajani, and D\. Lopez\-Paz\(2019\)Invariant risk minimization\.arXiv preprint arXiv:1907\.02893\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p1.1)\.
- \[2\]S\. Azzolin, A\. Longa, P\. Barbiero, P\. Lio, and A\. Passerini\(2023\)Global explainability of GNNs via logic combination of learned concepts\.InThe Eleventh International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[3\]S\. Azzolin, A\. Longa, S\. Teso, and A\. Passerini\(2025\)Reconsidering faithfulness in regular, self\-explainable and domain invariant gnns\.InInternational Conference on Learning Representations,Vol\.2025\.Cited by:[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[4\]S\. Azzolin, S\. Teso, B\. Lepri, A\. Passerini, and S\. Malhotra\(2026\)GNN explanations that do not explain and how to find them\.InThe Fourteenth International Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2608.14121#S2.p2.1),[§4](https://arxiv.org/html/2608.14121#S4.p2.1),[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[5\]V\. M\. De Luca, A\. Longa, P\. Lio, and A\. Passerini\(2025\)XAI\-drop: don’t use what you cannot explain\.InLearning on Graphs Conference,pp\. 16–1\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[6\]R\. Geirhos, J\. Jacobsen, C\. Michaelis, R\. Zemel, W\. Brendel, M\. Bethge, and F\. A\. Wichmann\(2020\)Shortcut learning in deep neural networks\.Nature Machine Intelligence2\(11\),pp\. 665–673\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p1.1),[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[7\]W\. Hamilton, Z\. Ying, and J\. Leskovec\(2017\)Inductive representation learning on large graphs\.Advances in neural information processing systems30\.Cited by:[§4](https://arxiv.org/html/2608.14121#S4.p3.1)\.
- \[8\]T\. N\. Kipf and M\. Welling\(2017\)Semi\-supervised classification with graph convolutional networks\.InInternational Conference on Learning Representations,Cited by:[§4](https://arxiv.org/html/2608.14121#S4.p3.1)\.
- \[9\]B\. Knyazev, G\. W\. Taylor, and M\. Amer\(2019\)Understanding attention and generalization in graph neural networks\.Advances in neural information processing systems32\.Cited by:[§4](https://arxiv.org/html/2608.14121#S4.p2.1)\.
- \[10\]D\. Krueger, E\. Caballero, J\. Jacobsen, A\. Zhang, J\. Binas, R\. L\. PRIOL, D\. Zhang, and A\. Courville\(2021\)Out\-of\-distribution generalization via risk extrapolation \(\{RE\}x\)\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p1.1)\.
- \[11\]S\. Lapuschkin, S\. Wäldchen, A\. Binder, G\. Montavon, W\. Samek, and K\. Müller\(2019\)Unmasking clever hans predictors and assessing what machines really learn\.Nature communications10\(1\),pp\. 1–8\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p1.1),[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[12\]D\. Li, Z\. Wang, Y\. Chen, R\. Jiang, W\. Ding, and M\. Okumura\(2024\)A survey on deep active learning: recent advances and new frontiers\.IEEE Transactions on Neural Networks and Learning Systems36\(4\),pp\. 5879–5899\.Cited by:[§3\.1](https://arxiv.org/html/2608.14121#S3.SS1.p3.1)\.
- \[13\]X\. Li, J\. Wang, and Z\. Yan\(2025\)Can graph neural networks be adequately explained? a survey\.ACM Comput\. Surv\.\.External Links:ISSN 0360\-0300,[Document](https://dx.doi.org/10.1145/3711122)Cited by:[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[14\]A\. Longa, S\. Azzolin, G\. Santin, G\. Cencetti, P\. Lio, B\. Lepri, and A\. Passerini\(2025\)Explaining the explainers in graph neural networks: a comparative study\.ACM Computing Surveys57\(5\),pp\. 1–37\.Cited by:[§3](https://arxiv.org/html/2608.14121#S3.p1.2)\.
- \[15\]S\. Miao, M\. Liu, and P\. Li\(2022\)Interpretable and generalizable graph learning via stochastic attention mechanism\.InInternational conference on machine learning,pp\. 15524–15543\.Cited by:[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[16\]A\. Pluska, P\. Welke, T\. Gärtner, and S\. MALHOTRA\(2024\)Logical distillation of graph neural networks\.InICML 2024 Workshop on Mechanistic Interpretability,Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p1.1),[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[17\]P\. E\. Pope, S\. Kolouri, M\. Rostami, C\. E\. Martin, and H\. Hoffmann\(2019\)Explainability methods for graph convolutional neural networks\.InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),Cited by:[§3](https://arxiv.org/html/2608.14121#S3.p1.2),[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[18\]A\. S\. Ross, M\. C\. Hughes, and F\. Doshi\-Velez\(2017\)Right for the right reasons: training differentiable models by constraining their explanations\.InProceedings of the 26th International Joint Conference on Artificial Intelligence,pp\. 2662–2670\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[19\]B\. Settles\(2009\)Active learning literature survey\.Cited by:[§3\.1](https://arxiv.org/html/2608.14121#S3.SS1.p3.1)\.
- \[20\]Z\. Song, Y\. Zhang, and I\. King\(2023\)No change, no gain: empowering graph neural networks with expected model change maximization for active learning\.Advances in neural information processing systems36,pp\. 47511–47526\.Cited by:[§3\.1](https://arxiv.org/html/2608.14121#S3.SS1.p3.1)\.
- \[21\]W\. Tai, T\. Zhong, G\. Trajcevski, and F\. Zhou\(2025\)Redundancy undermines the trustworthiness of self\-interpretable GNNs\.InForty\-second International Conference on Machine Learning,External Links:[Link](https://openreview.net/forum?id=hFvp9NYfY9)Cited by:[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[22\]W\. Tai, T\. Zhong, G\. Trajcevski, and F\. Zhou\(2026\)Self\-consistency improves the trustworthiness of self\-interpretable GNNs\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=hxGdAUn3sB)Cited by:[§5](https://arxiv.org/html/2608.14121#S5.p1.1)\.
- \[23\]S\. Teso, Ö\. Alkan, W\. Stammer, and E\. Daly\(2023\)Leveraging explanations in interactive machine learning: an overview\.Frontiers in Artificial Intelligence\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p2.1)\.
- \[24\]P\. Veličković, G\. Cucurull, A\. Casanova, A\. Romero, P\. Liò, and Y\. Bengio\(2018\)Graph attention networks\.InInternational Conference on Learning Representations,Cited by:[§4](https://arxiv.org/html/2608.14121#S4.p3.1)\.
- \[25\]K\. Xu, W\. Hu, J\. Leskovec, and S\. Jegelka\(2019\)How powerful are graph neural networks?\.InInternational Conference on Learning Representations,Cited by:[§4](https://arxiv.org/html/2608.14121#S4.p3.1)\.
- \[26\]Y\. Yang, E\. Gan, G\. K\. Dziugaite, and B\. Mirzasoleiman\(2024\)Identifying spurious biases early in training through the lens of simplicity bias\.InInternational conference on artificial intelligence and statistics,pp\. 2953–2961\.Cited by:[§2](https://arxiv.org/html/2608.14121#S2.p3.1),[2nd item](https://arxiv.org/html/2608.14121#S3.I2.i2.p1.1)\.
- \[27\]H\. Yuan, H\. Yu, S\. Gui, and S\. Ji\(2023\)Explainability in graph neural networks: a taxonomic survey\.IEEE Trans\. Pattern Anal\. Mach\. Intell\.,pp\. 5782–5799\.External Links:ISSN 0162\-8828,[Document](https://dx.doi.org/10.1109/TPAMI.2022.3204236)Cited by:[§3](https://arxiv.org/html/2608.14121#S3.p1.2)\.
- \[28\]D\. Zhang, S\. Betala, and C\. Agarwal\(2026\)Quantifying explanation quality in graph neural networks using out\-of\-distribution generalization\.arXiv preprint arXiv:2602\.07708\.Cited by:[§1](https://arxiv.org/html/2608.14121#S1.p3.1),[§3](https://arxiv.org/html/2608.14121#S3.p1.2),[§3](https://arxiv.org/html/2608.14121#S3.p1.3)\.
- \[29\]J\. Zhang, J\. Katz\-Samuels, and R\. Nowak\(2022\)Galaxy: graph\-based active learning at the extreme\.InInternational Conference on Machine Learning,pp\. 26223–26238\.Cited by:[§3\.1](https://arxiv.org/html/2608.14121#S3.SS1.p3.1)\.

## Appendix 0\.AGNN Details

In this section, we present the hyperparameters for different GNN architectures across both datasets in the passive supervision scenario\. Table[3](https://arxiv.org/html/2608.14121#Pt0.A1.T3)shows these values\. We set the batch size to1616and256256forER\-colorandCPatchMNIST, respectively, across all models and models\. We evaluate all methods in a wide range of learning rates:\(1​e−7,1​e−3\)\(1e\-7,1e\-3\)andλ\\lambdabetween11and10001000\. For a fair comparison, we fix the number of epochs for the passive methods and set a fixed value across the active query strategies\. As mentioned before, because early stopping can halt training before the effect of shortcuts becomes apparent, we deliberately do not use it\. We plan to release the full codebase to reproduce our experiments upon acceptance\.

Table 3:Hyperparameter settings for different GNN architectures for the passive supervision setup onER\-colorandCPatchMNIST\. “H” indicated the hidden dimension, “Aggr\.” indicated the aggregation function of the GNN, “Start E\.” indicates the starting epoch for training with both losses, and “E”\. indicated the total training epochs\.
## Appendix 0\.BAL Details

For active learning experiments, we setqqto1010for both datasets\. ForER\-color, we query 5 instances per round for 20 rounds and 30 epochs per round\. ForCPatchMNIST, we query 50 instances per round for 10 rounds and 30 epochs per round\. These settings hold for all models\. The learning rates andλ\\lambda\(varying in the same range as the passive experiments\) for each GNN, query strategy, and dataset combination are shown in Table[4](https://arxiv.org/html/2608.14121#Pt0.A2.T4)\.

Table 4:Training hyperparameters for different GNN architectures onER\-colorandCPatchMNIST\.