acceptodds
Under review as a conference paper at ICLR 2027

Graph Actor-Critic: Communication-Aware Message Passing in Graph Neural Networks

Abstract

Graph neural networks (GNNs) have shown remarkable success on homophilic graphs, where connected nodes tend to share similar features or labels. However, their performance degrades significantly on heterophilic graphs, where connected nodes often have different characteristics. In this work, we propose a novel reinforcement learning (RL) framework for adaptive graph rewiring that learns to dynamically modify graph structure to improve GNN performance on heterophilic graphs. We formulate graph rewiring as a Markov Decision Process (MDP) where an RL agent learns to select actions that optimize a reward function that encourages diverse node representations, thereby minimizing oversmoothing, all without requiring labeled data. We evaluate our method on diverse graph datasets and show that our RL-based rewiring approach, Graph Actor-Critic (GAC), significantly outperforms standard GNNs and existing rewiring methods on heterophilic graphs, while maintaining competitive performance on homophilic graphs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
controlled message passing, oversmoothing, graph rewiring

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_M1Z3OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Graph Actor-Critic (GAC), a reinforcement learning framework that rewires message passing to improve performance on heterophilic graphs. It casts each layer as a step of an MDP, in which a node-level "communication policy" decides whether each node sends, receives, both or neither. The policy acts over the original and k-nearest-neighbor candidate edges and is trained with a label-free reward that favors dissimilar neighbor embeddings.

Critical strengths

S1: I support the use of the heterophilic benchmark of Platonov et al. for node classification in place of older heterophilic datasets.

S2: The layer-wise visualizations on the toy graphs make the learned rewiring easy to inspect.

S3: The anonymized code is provided, which makes the results possible to check.

Critical weaknesses

W1: There are already many architectures designed for heterophily, and recent graph foundation models using standard message passing [1, 2, 3] also perform well on heterophilic datasets. I think the motivation needs to be stronger, as it is not clear what limitation of existing methods this paper addresses.

W2: I don't understand how the MDP formulation matches the actual method. The transitions seem to depend on the soft send/receive probabilities rather than the sampled actions. If so, the sampled actions do not affect the reward, and I don't see how the finite-action existence result applies. Please correct me if I'm missing something.

W3: The reward pushes connected node embeddings apart regardless of their labels. This might help on heterophilic graphs, but could hurt on homophilic ones. Why not learn the communication policy directly from the task loss, as in Co-GNN [4]?

W4: I think some important ablations are missing. In particular, how much of the improvement comes from the learned policy and how much from adding kNN edges? Also, the gain attributed to label-free rewiring is measured against GCN, rather than using the same backbone on the original graph.

W5: The main graph-level experiments use older datasets whose relevance has been questioned [5], and these are not specifically heterophilic benchmarks. I would like to see more evaluation on newer, larger graph-level benchmarks, such as [6, 7, 8].

References

[1] Fully-inductive Node Classification on Arbitrary Graphs, ICLR 2025.

[2] Equivariance Everywhere All at Once: A Recipe for Graph Foundation Models, NeurIPS 2025.

[3] GraphPFN: A Prior-Data Fitted Graph Foundation Model.

[4] Cooperative Graph Neural Networks, ICML 2024.

[5] Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks, ICML 2025.

[6] GraphBench: Next-generation graph learning benchmarking.

[7] GraphLand: Evaluating Graph Machine Learning Models on Diverse Industrial Data, NeurIPS 2025.

[8] OgBench: A Framework for Evaluating Graph Neural Networks on Omics Data.

Questions for authors

Q1: Do the sampled actions or the soft send and receive probabilities determine the next state? If the latter, how does GAC compare with optimizing the same reward by backpropagating through the gates?

Q2: Can you report, per dataset, what fraction of the pruned and created edges connect same-class nodes?

Q3: Several baseline results on the heterophilic benchmarks do not match the original sources and are mostly lower. Can the authors report the baselines with the numbers from their original papers, or retune them with the same budget as GAC, including tuned classic GNNs?

5 comments

We thank the reviewer for reviewing our paper and providing valuable feedback. We have gone over comments carefully and address the key concerns below-

W1: Existing architectures

The issue with existing methods is that nearly all methods including heterophily specific architectures and the newer graph foundation models run message passing on a fixed communication graph. At every layer, every node exchanges messages with it’s neighbor regardless of the representation/embedding quality. Hence, we can’t train deeper gnn models as the representation collapses (oversmoothing). Heterophily specific architecture changes how the messages are combined by applying spectral filters or higher neighborhood aggregation but it remains fixed across the layers which again leads to oversmoothing. In both cases, the question of who should communicate with at which layer is fixed. This is the key question GAC addresses. GAC allows nodes to decide which nodes to share messages, how much message to share and at which layer. The work done by [1] is the closest to our approach, however [1] is unable to create new edges and the communication is learned by downstream task signal which provably makes the model to learn shortcut and hence the same phenomena of representation collapse happens. In GAC, we learn the communication policy based on a reward designed to keep diversity while maintaining similarity to some extent which is proved by GAC’s performance on homophilic graphs. We have updated our motivation in the manuscript based on your feedback.

W2: MDP formulation matches..

MDP provides the base for our graph actor critic. Each node takes a soft send/receive action and their joint probability determines the transition to the next state. For example, for a given graph we have u-v pair, if node u’s send value is 1 and node v’s receive value is 0, then for the next state they won’t have any connectivity between them (drop edge from u→v). Similarly, if u’s receive value is 1 and v’s send value is 1 then for the next layer keep u←v edge (making it directional). The same idea applies to each node of the graph and we get a modified graph connectivity for the next layer. Each node’s action is determined by the reward to keep diversity in the node's representation. If neighbors' representation gets too similar, our reward penalizes each node’s action so that they don’t receive information from neighboring nodes.

W3: The reward pushes..

We have performed experiments on homophilic graphs and have shown that GAC performance doesn’t degrade on such graphs. Please check table 9 in the appendix.

Learning communication policy directly encourages the model to learn shortcut and we have to learn the policy for each task. Rather we show that it is possible to learn label free communication policy and apply the learned policy for downstream task such as classification, line prediction.

W4: Important ablations are missing..

To isolate the contribution of learned communication policy and topology augmentation with kNN neighbors, we perform ablation study on two datasets which are randomly chosen. We consider four configurations - 1. The backbone architecture (GCN with mean aggregation which is termed as mean gnn in literatures like [1] with the original graph, 2. Keep kNN edges only learned by the policy and drop existing edges from the original graph, 3. Remove kNN edges and keep the original edges learned by the policy, 4. Keep kNN edges and original edges with learned policy. The experiment configuration are kept the same and we used compute budge of 500 epochs.

ConfigurationRoman EmpireQuestions
mean-gnn (top_k=0)0.7666 ± 0.01430.7579 ± 0.0137
kNN Edges Only (top_k=8)0.7924 ± 0.00450.7122 ± 0.0231
reweighting Only (top_k=0, learned policy)0.8775 ± 0.00800.7741 ± 0.0125
full method (top_k=8, learned policy)0.8983 ± 0.00840.7876 ± 0.0238

From this ablation, it is clear that the learned communication policy is the primary contributor to performance improvements. Compared with the baseline (gcn with mean aggregation), actor critic policy-based reweighting improves the test metric from 0.7666 to 0.8775 on roman-empire and from 0.7579 to 0.7741 on questions datasets.

Q3: Baseline results on the heterophilic benchmarks

We ran all baselines using their official open-source implementations under the exact same experimental environment setup and compute budget (number of epochs, hyperparameter tuning range, and evaluation splits) to ensure a fair comparison. Under this uniform budget, we consistently reproduced the baseline performance numbers reported in our paper. For instance, methods like [1] also yield lower performance when strictly aligned with this standardized budget as shown in [2].

References:

[1] Cooperative Graph Neural Networks, Finkelshtein et. al, ICML 2024

[2] Cooperative sheaf neural networks. Ribeiro et.al. ICLR 2026

W5: The main graph-level experiments use older datasets..

We ran experiments on graphland dataset from pyg framework and report the following. In the original submission, please note that we have performed experiments on H2GB [1] (KDD'25) datasets where GAC achieves SOTA performance on that dataset (in the appendix table 11).

DatasetMetricGACmean-gnn
Tolokers-2AUC-ROC ↑0.8356 ± 0.02140.8061 ± 0.0154
Artnet-expAUC-ROC ↑0.7963 ± 0.02160.7800 ± 0.0354
City-reviewsAUC-ROC ↑0.9336 ± 0.03450.9122 ± 0.0154
Artnet-views (Result 1)MSE/MAE ↓0.5650 ± 0.01630.5789 ± 0.0121
City-roadsMSE/MAE ↓0.4121 ± 0.01420.4398 ± 0.0221
Artnet-views (Result 2)MSE/MAE ↓0.4826 ± 0.01820.5261 ± 0.0231
DatasetTaskNodesEdgesFeaturesTrain / Val / TestEdge Homophily
Tolokers-2AUC-ROC11,7581,038,000191,175 / 1,176 / 9,4070.5945
Artnet-expAUC-ROC50,405560,696755,040 / 5,041 / 40,3240.8317
City-reviewsAUC-ROC148,8012,330,83020469,439 / 34,720 / 34,7200.7392
Artnet-viewsMSE/MAE50,405560,696505,040 / 5,041 / 40,324—
City-roadsMSE/MAE57,073214,2086818,053 / 9,027 / 9,027—

We will run more experiments on the remaining datasets and other benchmarks you mentioned and will be available for the camera ready version.

[1] When Heterophily Meets Heterogeneity: Challenges and a New Large-Scale Graph Benchmark, Lin et al, KDD 2025

We again thank you for the review and feedback. We look forward to hearing back from you and answer any additional questions that you may have.

Best

The authors

Thank you for the rebuttal and new ablations.

Regarding W2/Q1, using the policy's continuous probabilities inside Eq. (4) makes state transitions depend on the policy itself rather than only the state and discrete action, so the standard finite-action MDP setup in Theorem 3.3 still does not hold, and the comparison against directly backpropagating the reward through the gates was not run.

Regarding Q2, Q3, and W5, Q2 was left unanswered, published Co-GNN results already match or beat GAC on the Platonov datasets (while GAC drops to 89.8 on Roman-Empire in the W4 ablation), and the added GraphLand tasks in W5 are node-level rather than graph-level benchmarks as mentioned in Q5.

Thus, the rebuttal partially address my concerns and I will keep my score of 2.

Reviewer_HcFAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Graph Actor-Critic (GAC), a reinforcement learning-based method for adaptive graph rewiring. The authors formulate layer-wise message passing in GNNs as a Markov Decision Process (MDP), where node-level policies determine whether nodes send or receive messages, and edge weights are adjusted according to the corresponding action probabilities. The method uses a label-free reward function that encourages high variance in edge differences and penalizes similarity between neighboring nodes to mitigate oversmoothing. The paper also provides theoretical analyses regarding the existence of an optimal policy and the pruning of edges that contribute to representation collapse. Experimental results show that GAC achieves strong performance on several heterophilic graph benchmarks while remaining competitive on homophilic graphs.

Critical strengths
  1. The paper formulates layer-wise message passing in GNNs as an MDP and controls information propagation through node-level send and receive actions. This formulation brings adaptive aggregation and graph rewiring into a unified framework. It offers a degree of novelty and has the potential to be extended to other GNN architectures.

  2. The experimental evaluation is relatively comprehensive. GAC achieves consistent improvements over existing baselines on multiple heterophilic node classification and graph classification datasets. These results also provide empirical evidence that label-free rewards can offer useful signals for learning graph structures.

Critical weaknesses
  1. There is an inconsistency between the action definition and the actual state transition in the MDP formulation. The paper defines four discrete communication actions for each node, but Equations (4) and (5) compute edge weights directly from the send and receive probabilities produced by the policy. It is unclear how the sampled discrete actions affect the state transition. If message passing is actually controlled by continuous probability-based gates, the applicability of the discrete-action policy gradient in Equation (21) needs further explanation. Moreover, since the state transition would depend on the policy distribution in this case, the MDP formulation and the theoretical guarantee in Theorem 3.3 also need to be reconsidered.

  2. The proof of Proposition 3.4 appears problematic. The authors claim that when neighboring nodes have similar representations, gradient ascent will eventually reduce the corresponding edge weights to zero. However, the edge-weighted objective differentiated in Equation (14) is not consistent with the original reward function defined in Equation (10). The proof also does not account for how changes in edge weights affect subsequent node representations. Furthermore, a negative local gradient does not necessarily guarantee that an edge weight will eventually converge to zero. Therefore, the current derivation does not seem sufficient to support the conclusion of Proposition 3.4.

  3. The node-level Critic lacks sufficient theoretical justification. Equation (19) approximates the joint action value as the sum of node-level Q-values, although communication actions between neighboring nodes are clearly coupled, since edge weights depend on the policies of both the sender and the receiver. In addition, Equation (20) uses the full graph-level reward to construct the TD target for every node, which may lead to double counting of rewards in Equation (19). Although the authors acknowledge the coupling between node actions in the limitations section, more explanation is needed regarding how this approximation affects value estimation and policy learning.

  4. The paper lacks an ablation study that isolates the contribution of reinforcement learning. Since GAC computes edge weights using continuous send and receive probabilities, it should also be possible to train the same gating network through direct gradient optimization. The current experiments demonstrate the effectiveness of the overall GAC framework, but they do not clearly distinguish whether the performance gains come from adaptive graph rewiring or from the Actor-Critic training procedure.

Questions for authors
  1. Could the authors clarify how node communication actions are executed? Are state transitions determined by sampled discrete actions or by continuous edge weights computed from policy probabilities? If the latter, how do the discrete actions affect the environment transitions, and how does the policy gradient in Equation (21) correspond to the actual optimization objective?

  2. Could the authors further explain the relationship between Equation (14) in Proposition 3.4 and the original reward function in Equation (10)? The current proof does not seem to directly establish that the edge weights will eventually converge to zero.

  3. Regarding the node-level Critic, could the authors explain how the value decomposition in Equation (19) is consistent with the TD target in Equation (20)? In particular, does using the full graph-level reward for every node lead to double counting?

  4. It would be helpful to include a comparison with direct gradient optimization of continuous gates to better establish the contribution of the Actor-Critic training procedure. I would also like to see an analysis of the model's sensitivity to the reward function hyperparameters, particularly the similarity penalty coefficient λ, along with additional ablation experiments.

2 comments

We thank the reviewer for their time to review our paper and provide feedback. Below we provide clarifications to the concerns you raised-

w1:

  1. How the sampled actions affect the transition and the trajectory?

We first clarify how the action determines edge weights- in GAC, each node takes a soft send/receive action and their joint probability determines the transition to the next state. At each layer ℓ\ell, every node u samples a discrete action au(ℓ)a_u^{(\ell)} from π(⋅∣s(ℓ))\pi(\cdot \mid s^{(\ell)}). The sampled action decides whether u sends and whether it receives or both or do nothing at all. The policy's probabilities [0,1] decide how much information to send or receive from neighboring nodes. The joint probability of each neighboring nodes action decide the edge weight i.e the connectivity strength between them. The edge weight is wu→v(ℓ+1)=αu(ℓ)βv(ℓ)w_{u\to v}^{(\ell+1)} = \alpha_u^{(\ell)} \beta_v^{(\ell)}, so a message flows from u to v only if u sends (α\alpha) and v receives (β\beta). The sampled actions therefore determine the connectivity of the next layer directly.

Because the graph at the next layer changes, the node embeddings change too. This affects the next set of candidate edges and the actions taken at the following layer. Different sampled actions therefore lead to different paths through the network.

  1. Why the policy gradient in Eq. (21) applies?

Since the sampled action decides which connections open or close, it has a direct effect on the next layer and on the reward. The advantage in Eq. (21) measures whether a node's choice turned out better or worse than the critic expected, so Eq. (21) is the standard actor-critic update for this choice.

The update is easy to read when sending and receiving are considered separately. We have added this breakdown to Section 3.3 adding further explanation to Eq 21:

  • If a node chose to send and the outcome was better than expected, it becomes more likely to send, and it sends more strongly.
  • If the outcome was worse than expected, it becomes less likely to send.
  • Receiving is updated in the same way.

An edge is reinforced only when its sender keeps sending and its receiver keeps receiving. This is how the policy learns to cut edges that make node representations collapse (Proposition 3.4). The entropy term keeps the agent from settling on one action too early. We have added further clarification after equation 21 in the revised manuscript. Please check.

  1. The MDP formulation and Theorem 3.3.

Following MDP definition, in GAC - tranisition from current state (i.e layer) to the next layer depends on the actions taken by the nodes. Given the current state of the graph and node actions, the next state is a deterministic function through one message passing step. Our lemma 3.1 holds true for this. Actions are bounded and the we can show that each gnn step is continuous and linearly bounded and the reward is continuous. Hence, backward induction over finite horizon can be dervied for GAC following [1]. We have updated the proof and added more clarity on it in the Appendix.

w2

We thank the reviewer for careful reading. We have read and verified the proof and respectfully believe that the proposition is correct as stated. The concerns may be arising from some points that are implicit in the original text and we are happy to clarify. Eq. (14) and Eq. (10) are consistent and they are not a different reward function. We have provided the detail calculation of this in Appendix A.2 where we decompose the variance term (Eq. (37) and write the reward function as

r(ℓ)=2−(2+λ)Sim(ℓ)−∥E(u,v)∼E[Δuv(ℓ)]∥22r^{(\ell)} = 2 - (2+\lambda)\mathrm{Sim}^{(\ell)} - \big\lVert \mathbb{E}_{(u,v)\sim\mathcal{E}}[\Delta_{uv}^{(\ell)}] \big\rVert_2^2

Here, the first two terms form an average the edges. Since the embeddings are normalized, Eq. (30) gives ∥Δuv∥22=2(1−hˉu⋅hˉv)\|\Delta_{uv}\|_2^2 = 2(1 - \bar{h}_u \cdot \bar{h}_v), so

2−(2+λ)Sim(ℓ)=E(u,v)∼E[∥Δuv(ℓ)∥22−λ(hˉu(ℓ)⋅hˉv(ℓ))],2 - (2+\lambda)\mathrm{Sim}^{(\ell)} = \mathbb{E}_{(u,v)\sim\mathcal{E}}\Big[\lVert \Delta_{uv}^{(\ell)}\rVert_2^2 - \lambda(\bar{h}_u^{(\ell)} \cdot \bar{h}_v^{(\ell)})\Big],

which is exactly the expected per-edge term differentiated in Eq. (14). Taking a look at the right hand side, it depends only on the cosine similarity of two connected nodes which lies [-1,1]. When similarity is -1, it equals 4 + λ\lambda > 0. When similarity is 1 where the two nodes have the same representation it equals λ\lambda < 0. Hence, for any λ>0\lambda > 0, the derivative is strictly negative whenever two connected nodes are similar than this threshold and the gradient ascent reduces the weight of their edge. This is the condition we described in proposition 3.4. In short, given actions, we evaluate the state by measuring the reward function which drives the policy optimization and proposition 3.4 is about communication suppression if it is harmful for efficient representation learning.

References:

[1] Reinforcement Learning: An Introduction, Sutton and Barton 1996, 2020

w3:

For GAC, we follow the standard design of actor-critic methods for cooperative reinforcement learning for multi agents, where agents share a common reward and each agent learns a value for its own action and thereby a shared policy is learned. We do not consider a critic over join action as it is not feasible since the action space is too large 4|V| and infeasible to track. Hence, we resort to approximation and shared node level critic (Please check [1] for details). Given the current state each node samples it's action independently making the policy factorized across nodes. Under this factorization, the gradient of the expected return with respect to the policy of node u is a function of the joint action value only via the average over other nodes' actions: Qu(s,au)=Ea−u∼π[Q(s,au,a−u)].Q_u(s, a_u) = \mathbb{E}_{a_{-u} \sim \pi}\big[Q(s, a_u, a_{-u})\big].

This is exactly the value learned by the node-level critic. The target of its training in Eq. (20) is the shared return observed after node u takes action au, with other nodes acting according to the current policy.

w4

To isolate the contribution of learned communication policy and topology augmentation with kNN neighbors, we perform ablation study on two datasets which are randomly chosen. We consider four configurations - 1. The backbone architecture (GCN with mean aggregation which is termed as mean gnn in literatures like [cite CoGNN paper]) with the original graph, 2. Keep kNN edges only learned by the policy and drop existing edges from the original graph, 3. Remove kNN edges and keep the original edges learned by the policy, 4. Keep kNN edges and original edges with learned policy. The experiment configuration are kept the same and we used compute budge of 500 epochs.

ConfigurationRoman EmpireQuestions
Baseline (top_k=0, no policy)0.7666 ± 0.01430.7579 ± 0.0137
kNN Edges Only (top_k=8, no policy)0.7924 ± 0.00450.7122 ± 0.0231
Reweighting Only (top_k=0, learned policy)0.8775 ± 0.00800.7741 ± 0.0125
Full Method (top_k=8, learned policy)0.8983 ± 0.00840.7876 ± 0.0238

From this ablation, it is clear that the learned communication policy is the primary contributor to performance improvements. Compared with the baseline (gcn with mean aggregation), actor critic policy-based reweighting improves the test metric from 0.7666 to 0.8775 on roman-empire and from 0.7579 to 0.7741 on questions datasets.

References:

[1] Value-Decomposition Networks For Cooperative Multi-Agent Learning, Sunehag et.al. 2018, AAMAS

We thank you again for your feedback and we look forward to hearing back from you and answer any questions that you may have.

Best

The authors

Reviewer_ZmmuOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Graph Actor-Critic (GAC), a reinforcement learning framework for adaptive message passing in GNNs, particularly targeting heterophilic graphs. The authors formulate layer-wise graph communication as a Markov Decision Process (MDP), where each node can select among four communication actions: send-and-receive, send-only, receive-only, and idle. An actor-critic architecture learns node-level communication policies using a label-free reward function designed to encourage representation diversity and reduce oversmoothing. The framework also introduces potential new edges through kNN search in embedding space.

Critical strengths
  • Modeling node-level communication as a sequential decision-making process offers an interesting perspective on graph representation learning.
  • The proposed reward function does not directly require node labels, making the approach potentially applicable to settings with limited supervision. The two-stage training procedure, in which the communication policy is learned before supervised task training, is an appealing design choice.
Critical weaknesses
  • The paper does not clearly explain whether the sampled discrete actions actually determine the graph transitions, or whether message passing is performed directly using the continuous marginal probabilities.
  • GAC combines several components: actor-critic optimization, learned send/receive gates, embedding-based kNN edge construction, and a diversity-similarity reward; but it does not provide a systematic ablation study.
Questions for authors

See weaknesses,

1 comment

We thank the reviewer for their valuable time and providing feedback for our paper. Below we address the concerns raised by the reviewer-

w1: The paper does not clearly explain whether the sampled

The graph transition is determined by the node’s action which constitutes the graph connectivity for the next state. Let’s consider two nodes, u and v. Each of them can take four different actions. Let’s consider different scenarios for node u and v where each node takes two actions send and receive. Based on this, we can define the following scenarios-

  1. Send only : u → v (directed)
  2. Receive only: u ← v (directed)
  3. Send and receive: u ← → v (undirected)
  4. Idle: u –x– v (no information shared)

For example, for a given graph we have u-v pair, if node u’s send value is 0.5 and node v’s receive value is 1, then for the next state the connectivity strength between them is 0.5x1 = 0.5 i.e v receives the full message from u. Similarly, if u’s receive value is 1 and v’s send value is 1 then for the next layer keep u←v edge (making it directional and keep the original connectivity). Now consider all the nodes and edges in the graph, the entire graph will be transformed into a new state. We perform message passing on this new graph and evaluate the connectivity based on reward feedback. We have added more clarification on the transition dynamics of GAC in section 3.1, please check.

w2 GAC combines several components: actor-critic optimization...

To isolate the contribution of learned communication policy and topology augmentation with kNN neighbors, we perform ablation study on two datasets which are randomly chosen. We consider four configurations - 1. The backbone architecture (GCN with mean aggregation which is termed as mean gnn in literatures like [cite CoGNN paper]) with the original graph, 2. Keep kNN edges only learned by the policy and drop existing edges from the original graph, 3. Remove kNN edges and keep the original edges learned by the policy, 4. Keep kNN edges and original edges with learned policy. The experiment configuration are kept the same and we used compute budge of 500 epochs.

ConfigurationRoman EmpireQuestions
mean-gnn (top_k=0)0.7666 ± 0.01430.7579 ± 0.0137
kNN Edges Only (top_k=8)0.7924 ± 0.00450.7122 ± 0.0231
reweighting Only (top_k=0, learned policy)0.8775 ± 0.00800.7741 ± 0.0125
full method (top_k=8, learned policy)0.8983 ± 0.00840.7876 ± 0.0238

From this ablation, it is clear that the learned communication policy is the primary contributor to performance improvements. Compared with the baseline (gcn with mean aggregation), actor critic policy-based reweighting improves the test metric from 0.7666 to 0.8775 on roman-empire and from 0.7579 to 0.7741 on questions datasets.

Again, thank you for your feedback. We look forward to hearing back from you and clarify any concerns you may have.

Best

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.