acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Semantic Evolution-Driven Reinforced Denoising on Dynamic Graphs

Abstract

Dynamic graph learning relies on capturing complex topological and temporal evolutions. However, most models extract historical memory using a rigid exponential time-decay paradigm, struggling to model long-range dependencies and inducing severe extraction bias. This bias prevents models from reliably assessing incoming interactions, leading to indiscriminate aggregation that easily incorporates noisy links and propagates errors across nodes. To address this, we propose ATD-DyG, an adaptive evolution and reinforcement denoising framework. First, an adaptive semantic evolution kernel dynamically fits evolutionary patterns to reconstruct high-fidelity representations, eliminating temporal extraction bias. Using these representations as reliable decision priors, a reinforcement learning agent evaluates incoming interactions and proactively truncates high-risk noisy links. This prior-guided topological intervention shifts the paradigm from “passive absorption” to “proactive blocking”, effectively halting noise propagation and enhancing model robustness. Experiments on 7 benchmark datasets demonstrate that ATD-DyG achieves advanced predictive performance and remarkable noise resistance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic graph neural network, semantic evolution, reinforcement denoising, time‑decay, temporal dependency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_N2SVOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ATD-DyG, a continuous-time dynamic graph learning framework that aims to address temporal extraction bias and structural noise propagation. The method consists of two main components. First, an adaptive semantic evolution kernel based on a truncated Mittag-Leffler expansion adjusts temporal memory according to semantic similarity, with the goal of preserving long-range but relevant historical interactions while rapidly forgetting semantically inconsistent ones. Second, the paper formulates neighborhood denoising as a sequential subset-selection problem and uses policy-based reinforcement learning to make hard edge-retention decisions. The step-wise reward is defined as the marginal improvement in the log-likelihood of the target link, so that the cumulative reward is aligned with the downstream link-prediction objective. The method is evaluated against 11 CTDG baselines on seven datasets under both transductive and inductive settings, together with robustness, ablation, efficiency, and stability analyses.

Critical strengths
  1. Coherent problem formulation and method design. The paper provides a clear motivation for coupling temporal modeling with structural denoising. In particular, the adaptive semantic evolution mechanism attempts to distinguish semantic relevance from physical recency, while the subsequent reinforcement-learning module uses the resulting representations as decision priors for hard neighborhood selection. This makes the overall framework conceptually coherent rather than a simple combination of unrelated components.
  2. Strong empirical coverage and well-motivated reward design. The paper evaluates the proposed model against 11 baselines across seven datasets under both transductive and inductive settings, where ATD-DyG obtains strong average rankings. In addition, defining the reinforcement-learning reward as the marginal improvement in target-link log-likelihood leads to a telescoping cumulative reward that is directly connected to the final prediction objective, which is a meaningful design choice.
Critical weaknesses
  1. There is a substantial mismatch between the theoretical analysis of the adaptive semantic evolution kernel and its actual implementation. The proposed model uses a finite truncated Mittag-Leffler expansion, and the experiments ultimately adopt K=20 for computational reasons. However, the theoretical argument supporting the transition between exponential forgetting and long-tail power-law memory relies on the limiting behavior of the full Mittag-Leffler function, including the limit as K approaches infinity. Therefore, the current analysis does not establish that the implemented finite-K kernel actually exhibits the asymptotic memory behavior that forms a central motivation of the method. This issue is particularly important because the claimed long-range memory property is one of the main technical justifications for the adaptive evolution module.
  2. The current empirical analysis does not sufficiently isolate or validate the paper's central design choices. The main RL ablations compare the full method with Select-All, Select-None, and Select-Random variants. These are relatively weak controls and do not establish whether policy-based reinforcement learning is necessary compared with stronger deterministic or differentiable alternatives such as learned gating, threshold-based pruning, or top-k selection. Similarly, replacing the Mittag-Leffler kernel with exponential decay does not disentangle the contribution of fractional memory from the contribution of semantic adaptation of the fractional order. The robustness claim is also broader than the current evidence: controlled structural-noise injection is mainly demonstrated on Enron, despite denoising being a central contribution of the paper. Moreover, the exact construction of the injected structural noise should be described more explicitly to ensure reproducibility.
Questions for authors
  1. Can the authors provide a theoretical error bound or a finite-K analysis showing that the truncated Mittag-Leffler kernel used in the actual implementation, particularly with K=20, preserves the claimed long-tail memory behavior over the temporal range encountered in the datasets? Without such an analysis, it is unclear whether the asymptotic properties derived for the limiting function apply to the evaluated model.
  2. Can the authors provide stronger ablations that separately evaluate fixed fractional order versus semantic-adaptive fractional order, and compare the RL-based hard selection mechanism with competitive non-RL pruning strategies such as learned gating, threshold-based selection, or deterministic top-k selection?
  3. Could the authors clarify the exact structural-noise generation protocol and evaluate robustness on additional datasets or perturbation types? Since robustness and proactive denoising are central claims of the paper, broader controlled evaluation would materially affect my assessment.
Reviewer_L7QGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper points out that most existing dynamic graph models use rigid exponential time decay patterns to extract historical memory, which makes it hard to model long-range dependencies and can lead to serious extraction bias. This often causes them to forget useful long-term information and accumulate noisy interactions, affecting prediction performance and robustness. To address this, it proposes ATD-DyG, which dynamically adjusts the time decay weights of historical interactions through an adaptive semantic evolution kernel to capture long-term evolving dependencies. At the same time, it introduces a reinforcement learning-driven neighbor selection mechanism to actively filter and cut off potential noisy interactions based on downstream prediction benefits. This framework combines adaptive time modeling with proactive structural denoising, aiming to improve the accuracy and robustness of dynamic graph representations, and its prediction performance and noise resistance are validated on seven dynamic graph benchmark datasets.

Critical strengths

This paper proposes a complementary framework for temporal modeling and proactive denoising. The paper tackles two issues in dynamic graphs: insufficient long-term dependency modeling and the propagation of structural noise. It introduces a unified framework that combines adaptive Mittag-Leffler memory kernels with reinforcement learning-based neighbor selection. This design not only focuses on prediction accuracy but also aims to improve the robustness of dynamic graph representations from the perspectives of temporal dependency modeling and neighborhood structure optimization. It has some methodological innovation and practical significance, offering an insightful research direction for dynamic graph representation learning.

Critical weaknesses

1.The experimental results are unreliable. In Table 4, for the UCI and MOOC datasets under the Inductive setting, the mean AP and standard deviation of all 8 baseline methods (JODIE, DyRep, TGAT, TGN, CAWN, TCL, GraphMixer, and DyGFormer) are exactly the same. Considering that the structural features and interaction patterns of the two datasets are different, such extensive result repetition is questionable. Additionally, the results under the same experimental settings in Tables 1, 3, and 7 are inconsistent. These anomalies might be due to data duplication or typesetting errors, but the authors haven’t explained them, which undermines the rigor and reproducibility of the experiments and casts doubt on the claimed performance advantages and related conclusions of the paper.

2.Lack of direct evidence for the effectiveness of the core denoising mechanism. The paper claims that actively identifying and cutting off noisy interactions is its main contribution, but its reinforcement learning reward function is mainly optimized based on the log-likelihood gain of target link prediction, rather than directly measuring whether an interaction is structural noise. Therefore, the neighbors filtered by the model aren't necessarily truly noise-free interactions. Additionally, the paper's robustness experiments are only conducted on the Enron dataset and mainly evaluate denoising effects through downstream prediction performance, without directly analyzing metrics like noise edge identification accuracy or noise retention rate. So, while the existing experiments show that the method can improve prediction performance, they aren't enough to support the core claim that it can effectively identify noise and block its propagation.

Questions for authors

1.This paper attributes the main limitations of existing dynamic graph models as their reliance on monotonically exponential time decay, but methods like DyGFormer don’t actually use a fixed exponential decay mechanism. Figure 1 only shows the deviation between the structural correlation of historical neighbors and the exponential decay curve for randomly selected target nodes, lacking a statistically representative systematic analysis, and it can’t prove that existing models actually have such time modeling bias. With this core research motivation not fully validated, how can the authors prove that the problem they raise really exists and is general, and use that to justify the necessity of the adaptive temporal evolution kernel and the main contribution of the paper?

2.The proof of the long-tail decay property of the Mittag-Leffler memory kernel in this paper is based on an infinite series, but in practice a finite truncation (K=20) is used. Is there a strict theoretical guarantee or direct evidence showing that the truncated kernel still maintains the claimed decay property? Also, Figure 7 suggests that increasing K can further improve prediction performance. How do the authors demonstrate that K=20 is sufficient to accurately approximate the original memory kernel, rather than it just being an empirical trade-off for computational cost?

3.This paper takes active structure denoising as its main contribution, but the reinforcement learning reward is only determined by the link prediction gain, which doesn't directly correspond to noise identification. Without results on noisy edge recognition and a thorough comparison with learnable selection strategies, how can we prove that the model really achieves structural denoising, rather than just task-driven neighbor selection? Where does the unique benefit of the reinforcement learning mechanism show up?

4.There are inconsistencies in the results under the same experimental setup in Table 1, Table 3, and Table 7, and in Table 4, some baseline results for different datasets are exactly the same. Are these differences due to typos, changes in the experimental setup, or issues with result calculation?

Reviewer_9tJBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ATD-DyG, a framework for continuous-time dynamic graph learning that combines adaptive temporal modeling with reinforcement learning-based neighborhood denoising. It introduces an adaptive Mittag-Leffler memory kernel to capture long-range temporal dependencies. It also formulates neighbor selection as a sequential decision process, rewarding decisions that improve link prediction. Experiments on seven benchmarks under transductive and inductive settings report competitive link prediction performance, while additional experiments on Enron evaluate robustness to structural noise.

Critical strengths

The adaptive memory kernel adjusts the weights of historical interactions according to semantic similarity, helping retain older interactions that remain relevant.

The reward formulation has a clear connection to the prediction objective, providing a concrete justification for the objective used to train the selection policy.

Critical weaknesses

The paper lacks the required AI use statement.

There is an inconsistency between the implemented memory kernel and its theoretical justification. Equation (2) defines a finite series, and Appendix B.7 specifies K = 20, whereas Appendix A.5 derives the decay properties of the infinite Mittag-Leffler series. For fixed K ≥ 1 and α > 0, the highest term of the finite series grows in magnitude as Δt becomes large, rather than exhibiting the claimed asymptotic decay.

The experiments do not sufficiently establish the claimed mechanism of identifying noisy interactions and blocking their propagation.

Questions for authors

Could you clarify how the decay properties derived in Appendix A.5 apply to the finite series in Equation (2), given that the proof considers the infinite Mittag-Leffler series while the implementation uses K = 20?

Are rejected interactions excluded only from the final aggregation, or also from subsequent historical state updates?

Under the same candidate set and a matched neighbor retention rate, how does the learned policy compare with selecting neighbors by the highest attention scores or an attention threshold?

Please visualize the temporal weights assigned by the adaptive memory kernel to interactions with different semantic similarities, alongside those from exponential decay. This would help verify whether the implemented kernel preserves relevant historical interactions over longer time intervals.

Please visualize the selection probabilities and final retention decisions for injected noisy edges and original edges across different temporal distances. The current retention plots do not show whether the agent distinguishes noise from informative interactions.

Could you provide representative selection trajectories showing the neighbors retained or discarded, the reward at each step, and the corresponding changes in prediction confidence?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.