acceptodds
Withdrawn from ICLR 2027

A Dynamic Graph Learning Method with Backward Knowledge Feedback via Parameter Expansion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dynamic graphs play a crucial role in modeling temporal interaction systems in the real world. Current dynamic graph learning methods suffer from limited representation capability due to the following issues: (1) fixed model parameter capacity leads to knowledge squeezing when continuously learning new graph knowledge; (2) such methods adhere to the fundamental paradigm of leveraging historical graph knowledge to learn new graph knowledge, neglecting the potential of backward knowledge feedback for enhancing the learning of historical graph knowledge. To address these issues, we propose a novel Dynamic graph learning method with Backward Knowledge Feedback via Parameter Expansion (DBKF-PE), which achieves effective representation of dynamic graphs through a bidirectional knowledge circulation structure. The circulation structure consists of forward knowledge injection and backward knowledge feedback. Specifically, the forward knowledge injection incorporates a new graph-aware parameter expansion strategy. This strategy expands new parameters to learn new graph knowledge while freezing old parameters, thereby preventing conflicts between new and historical graph knowledge in the parameter space. Then, the backward knowledge feedback is modeled as a meta-learning optimization objective, which not only alleviates catastrophic forgetting through the proposed backward knowledge feedback mechanism but also further enhances the performance of the new model on historical graphs. Finally, extensive experiments on six datasets validate that DBKF-PE outperforms the state-of-the-art baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic graph learning, Backward knowledge feedback, Parameter expansion, Catastrophic forgetting

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_R5tnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a Backward Knowledge Feedback via Parameter Expansion (DBKF-PE) method for dynamic graph learning. DBKF-PE incorporates parameter expansion to alleviate knowledge squeezing and backward knowledge feedback to ensure model performance on historical graphs. The proposed method is evaluated on six public datasets and shows improved results.

Critical strengths
  1. The idea of using newly learned graph knowledge to improve performance on historical graphs is interesting and worth exploring.
  2. The experiments cover multiple datasets and downstream tasks, with generally competitive results against existing methods.
Critical weaknesses
  1. Eq. (1) is confusing. If GtrG_t^r represents the unchanged part of GtG_t relative to Gt−1G_{t-1}, then Gt−1rG_{t-1}^r should be defined relative to Gt−2G_{t-2}. It is unclear why Gt−1rG_{t-1}^r can be used to construct GtG_t. In addition, where should the edges between GtrG^{r}_t and GteG^{e}_t be?
  2. The motivation behind Eq. (2) is unclear. Why does the model minimize the loss on the entire graph while maximizing the loss on the changed subgraph? It is also unclear why the authors cannot simply optimize the GNN on the unchanged subgraph. It is unclear what parameters are being expanded; are they additional GNN layers, higher-dimensional vectors, or something else? The paper needs to define the expansion clearly.
  3. The parameter freezing strategy seems inconsistent. Section 3.2 claims that freezing historical parameters prevents catastrophic forgetting, but Section 3.3 allows these parameters to be updated. This raises questions about how the method actually prevents forgetting.
  4. The paper is not well-written and is very difficult to follow.
Questions for authors
  1. What is the definition of GtrG_{t}^r?
  2. What is the rationale for Eq. (2)?
  3. What are the expanded model parameters?
  4. Can your method deal with node/edge deletion?
Reviewer_oeTZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DBKF-PE, a discrete-time dynamic graph learning framework designed to address knowledge squeezing and improve historical graph representations. The framework consists of forward knowledge injection and backward knowledge feedback. Forward injection employs graph partitioning and parameter expansion to learn evolving graph information while preserving historical knowledge. A temporal guidance mechanism captures dependencies between consecutive snapshots through an auxiliary prediction task. Backward knowledge feedback introduces a meta-learning-inspired objective to improve model performance on historical snapshots. The method is evaluated on node classification and community detection tasks across six benchmark datasets, with comparisons against existing dynamic graph learning and continual learning approaches.

Critical strengths

The paper explores backward knowledge feedback as an extension of conventional continual graph learning, shifting attention from merely preventing catastrophic forgetting to improving historical representations. The proposed framework combines graph-aware parameter expansion, temporal guidance, and historical performance optimization within a unified architecture. Its experimental evaluation covers two downstream tasks and six datasets, with comparisons against a broad set of baselines. The component-level ablations and additional parameter sensitivity experiments further support the empirical investigation of the proposed method.

Critical weaknesses

W1. The comparison between CTDG and DTDG methods is overly generalized. Section 1 claims that CTDG methods incur higher computational complexity, whereas DTDG methods are more efficient. However, event-driven CTDG architectures can efficiently process sparse temporal interactions without repeatedly constructing full graph snapshots. The computational trade-offs should be discussed more carefully rather than presenting DTDG as generally more efficient.

W2. The motivation overlooks memory-based dynamic graph learning methods. Section 1 attributes knowledge squeezing primarily to fixed model parameter capacity. However, memory-based dynamic GNNs can maintain evolving temporal states without explicitly expanding model parameters. The paper should clarify how its proposed approach differs from such architectures and why parameter expansion offers additional benefits.

W3. The backward knowledge feedback concept lacks a sufficiently concrete explanation. Section 1 and Figure 1 introduce bidirectional knowledge circulation but do not clearly illustrate how newly acquired knowledge improves historical graph representations. The distinction between backward feedback, conventional continual learning, and retrospective optimization remains unclear. A concrete example would strengthen the motivation.

W4. The graph-aware parameter expansion mechanism requires further clarification. Section 3.2 states that rational subgraphs can change across snapshots, but does not explain how their corresponding frozen parameters are identified and reused. Moreover, Equation (2) subtracts the environmental subgraph loss from the full-graph loss, potentially introducing conflicting optimization objectives. The parameter-mapping strategy and mathematical justification should be clarified.

W5. The experimental datasets provide limited evidence of generalization to realistic dynamic graphs. Section 4.1 evaluates six datasets, several of which originate from static graph benchmarks. The construction of temporal snapshots and the extent to which they reflect realistic graph evolution remain insufficiently documented. Additional naturally evolving datasets would strengthen the evaluation.

Questions for authors

refer to the weakness

Reviewer_yf1ROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DBKF-PE, a dynamic graph learning method for discrete-time dynamic graphs. It partitions each snapshot into a rational subgraph and an environmental subgraph, uses a graph-aware parameter expansion strategy to learn new graph knowledge while freezing historical parameters, adds a temporal guidance mechanism to capture temporal relations between consecutive snapshots, and introduces a backward knowledge feedback mechanism (BKFM) based on meta-learning to feed new-graph knowledge back to historical graphs. The stated goal is to move from unidirectional knowledge flow to a “bidirectional knowledge circulation” paradigm. The method is evaluated on six benchmark datasets for node classification and community detection, with comparisons against continual-learning and DGL baselines, ablation studies, backward-feedback experiments, and appendix experiments on parameter expansion, meta-loss weight, and model distillation.

Critical strengths

1、The paper targets a relevant limitation of existing DGL methods: unidirectional knowledge flow and weak performance on historical graphs. This is consequential because if backward knowledge feedback were properly formulated and validated, it could meaningfully improve model reusability and historical-graph performance. The motivation is therefore stronger than a generic incremental-learning framing. 2、The proposed pipeline combines parameter expansion, temporal auxiliary learning, and meta-learning in a coherent high-level structure, with experiments on multiple datasets and tasks. This is useful as a starting point and provides some empirical signal. However, the current evidence is not sufficient to establish that the central mechanism works as claimed.

Critical weaknesses
  1. The core BKFM and the claimed “bidirectional knowledge circulation” are not sufficiently specified or validated. The paper does not give a complete meta-learning objective, meta-gradient computation, or precise update procedure. It is also unclear whether historical graph representations are genuinely updated/enhanced or whether the new model is simply optimized to fit historical snapshots better. This is consequential because backward knowledge feedback is the paper’s main claimed contribution; without a clear mechanism and controlled evidence, the central claim is not established.

  2. The temporal guidance mechanism may have an information-availability or label-leakage problem, and the empirical comparison does not adequately control for model capacity. Equations (4)–(5) use ΔAt+1\Delta \mathbf{A}_{t+1} and ΔXt+1\Delta \mathbf{X}_{t+1} to construct temporal labels. The paper must clarify whether this information is available at inference time and, if so, why this is not leakage. In addition, parameter expansion increases capacity; the reported gains may partly reflect larger parameter budgets rather than the proposed mechanisms. This is consequential because both issues directly affect the validity of the main empirical conclusions.

Questions for authors
  1. Please provide the complete BKFM formulation: the exact meta-objective, the meta-gradient update, how the meta-loss is combined with the main loss, and which parameters are updated at each stage. Does θtr\theta_t^r remain frozen throughout, and how exactly are historical graph representations enhanced?

  2. At inference time, are ΔAt+1\Delta \mathbf{A}_{t+1} and ΔXt+1\Delta \mathbf{X}_{t+1} available? If not, how is the temporal auxiliary task used outside training? If yes, why does constructing temporal labels from t+1t+1 changes not constitute label leakage?

  3. How do you control for parameter count, training time, and memory cost relative to baselines? Please provide matched-capacity ablations or a clear parameter/efficiency comparison to substantiate that gains come from the proposed mechanisms rather than increased capacity.

  4. Are the reported improvements statistically significant across multiple runs? Please report mean and standard deviation, data splits, and hyperparameter-tuning protocols for all baselines.

  5. The introduction motivates node additions and deletions, but the method section focuses mainly on edge and attribute changes. How does DBKF-PE handle node additions/deletions in the graph partitioning, parameter expansion, and temporal guidance mechanism?

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

Related papers

Discussion (0)

Sign in to comment.