acceptodds
Under review as a conference paper at ICLR 2027

BayesMem-CD: Learning Stable Dynamic Graph Representations with Bayesian Temporal Memory

Abstract

Dynamic community detection is confronted with a fundamental ambiguity: an observed structural change may be indicative of genuine community evolution or merely transient graph noise. This issue is typically resolved by existing methods through deterministic temporal smoothing or snapshot-wise inference, which can either suppress meaningful structural transitions or generate unstable assignments. Additionally, deterministic temporal representations offer restricted insight into the periods during which historical evidence is uncertain. We introduce BayesMem-CD, an uncertainty-aware framework that is based on Bayesian temporal memory. This framework conceptualizes community evolution as a probabilistic temporal inference, rather than a series of deterministic clustering decisions. BayesMem-CD quantifies epistemic uncertainty and regulates temporal adaptation based on the reliability of observed structure by maintaining long-range structural evidence through graph memory and representing community assignments probabilistically. Spurious assignment changes are further discouraged by short-term temporal regularization, which does not preclude persistent evidence from driving genuine community transitions. Experiments conducted on synthetic and real-world dynamic networks have shown that BayesMem-CD is capable of achieving robust community recovery and temporal stability, as well as better-calibrated uncertainty under structural perturbations. The complementary benefits of probabilistic inference, temporal memory, and short-term consistency are further demonstrated by ablation and sensitivity analyzes. These findings indicate that a principled approach to distinguishing structural evolution from transitory variation in dynamic graphs is provided by explicitly modeling uncertainty in temporal evidence.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic Community Detection, Temporal Graph Learning, Graph Neural Networks, Bayesian Inference, Uncertainty Quantification, Temporal Memory, Probabilistic Community Detection, Dynamic Graphs

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7AJWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a BayesMem-CD, an uncertainty aware framework that presents temporal graph clustering as a Bayesian temporal inference problem. A graph memory based mechanism is applied to model the long range interactions of different nodes using a recurrence mechanism. Moreover, a short-term temporal regularization is applied to tackle spurious cluster assignment. The cluster assignment function is treated as a variational inference problem whereas the function is learned by using reparameterization trick. The proposed approach is experimented on both synthetic and real world datasets showing the impact of modeling uncertainty in dynamic graph clustering.

Critical strengths
  1. Quantifying the uncertainty in cluster assignment for dynamic graph is well motivated research problem and the paper presents it well.
  2. Experiment is done on both synthetic and real world dynamic graph datasets.
Critical weaknesses
  1. The paper is proposing BayesMem-CD, a bayesian temporal memory based community detection, however there’s no Bayesian part in the temporal memory modeling part. The name itself is confusing. Graph memory is updated by using well known gating mechanism that recurrently updates the node states over time. Instead uncertainty is modeled in the cluster assignment function using variational inference with reparameterization trick.

  2. The idea of the proposed approach is written on a high level and not justified (in the supplement materials too). Notations are not well defined and there are some missing notations. Detail ablation is missing for different loss component.

Questions for authors
  1. Can you explain how the reconstruction loss works?
  2. Given the community structure how we can reconstruct the graph topology?
  3. For learning the cluster assignment distribution, what is the reason for choosing a Gaussian distribution instead of categorical distribution?
  4. What is impact of number of samples in monte carlo averaging?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 7AJW

I appreciate your observation that the title and method description obfuscate the distinction between Bayesian community inference and temporal memory.

  1. Which component of the model is Bayesian? The graph memory is a deterministic, exponentially weighted update of fused snapshot representations. Over the weights and biases of the community assignment head, the Gaussian prior and variational posterior are defined. The memory-informed representation is received by that head, which then generates a distribution of assignments. We would refer to the method as Bayesian community inference with temporal graph memory and incorporate this terminology into the title, abstract, and method.
  2. Loss ablations and method detail Before presenting the objective, we would establish the observed edge-pair target, the posterior average, and the per-draw assignment probabilities. Additionally, we would differentiate between the temporal KL between consecutive predictions and the posterior-to-prior KL. Component-level evidence is presented in Table 2. The DSBM ECE is increased from 0.052 to 0.073 by removing the Bayesian component, while the Dynamic LFR CSI is reduced from 0.829 to 0.758 by removing the temporal KL. We would report matched runs that separately remove the auxiliary reconstruction loss, the variational term, the temporal KL, and the memory-alignment loss, while maintaining the relevant architecture and training settings, in order to isolate the loss terms.
  3. What is the method for calculating the reconstruction loss? We multiply the posterior-averaged membership probabilities of the two nodes for each community in a node pair and then sum the products across all K communities. In the main text, Eq. (17) denotes the dense squared error between the soft co-membership matrix and the adjacency matrix. The sparse implementation in supplementary Eq. (23) assesses binary cross-entropy on sampled non-edges and observed edges. The loss increases the score of observed edges and decreases the score of sampled non-edges. It is possible to calculate it pair by pair without the need to construct an N-by-N matrix.
  4. To what extent can the community structure be used to reconstruct the graph topology? The dot product is the likelihood that two memberships from the predicted rows that were independently selected agree. Assuming assortative structure, it is utilized as an edge score. Any third node is predicted to have identical affinities to two nodes with identical membership vectors. Consequently, the decoder captures connectivity at the community level, whereas the observed graph may contain individual edges and non-edges that the membership vectors are unable to differentiate.
  5. What is the rationale for employing a Gaussian distribution? The posterior is approximated by the Gaussian over the continuous weights and biases of the assignment head. It provides reparameterized sampling and a closed-form KL with the Gaussian prior. Under the condition of a parameter draw, the softmax output parameterizes a categorical distribution over the K possible memberships. In the head parameters, posterior averaging is used to integrate over uncertainty.
  6. What is the impact of the Monte Carlo sample count? Modularity was 0.506 and CSI was 0.762 with a single posterior draw. These values increased to 0.524 and 0.789 after three draws. They achieved 0.536 and 0.806 after five draws. The modularity was 0.534 and the CSI was 0.812 after ten draws.

In this sweep, both metrics were enhanced by increasing the sample count from one to five. The modularity was marginally reduced, and 0.006 CSI was added when the value was raised to ten. Five draws are the default. Monte Carlo variability is diminished by the addition of more draws, necessitating additional assignment-head evaluations. The variance estimator in Eq. (14) is identically zero with a single draw, as there is only one posterior prediction. The definition of predictive entropy remains unchanged.

Thank you to the authors for their response and providing clarification. I have carefully read the response and still believe that the manuscript requires to provide more theoretical clarification and experimental findings.

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

This paper presents BayesMem-CD, a Bayesian temporal memory framework for dynamic community detection designed to disambiguate genuine community evolution from transient structural perturbations. The method integrates a graph encoder, recurrent graph memory for historical structural retention, a variational Bayesian community predictor, and a KL-divergence-based short-term temporal regularizer, with predictive entropy and posterior predictive variance as uncertainty metrics. However, the current submission raises substantial concerns regarding methodological novelty, the justification of the Bayesian formulation, the validity of the temporal stability objective, and the rigor of experimental evidence. Most critically, the reported results do not convincingly establish that the proposed framework delivers a principled advantage over simpler deterministic temporal models.

Critical strengths
  1. Relevant and meaningful research problem. The paper addresses the difficulty of distinguishing transient structural noise from genuine community evolution. Explicitly modeling assignment uncertainty is potentially useful when community boundaries are ambiguous or network structure changes rapidly.

  2. Probabilistic predictions offer potentially useful information. The use of posterior sampling, predictive entropy, and posterior predictive variance provides a richer output than deterministic community labels. This could support downstream applications that need to assess the reliability of community assignments.

Critical weaknesses
  1. The Bayesian Learning Objective Is Underspecified. Although the supplementary material provides additional details on the structural likelihood and variational inference procedure, the relationship between these components in the overall training objective remains insufficiently clear. In particular, both the structural reconstruction loss and the likelihood term in the variational objective appear to be derived from the observed graph structure. The paper does not adequately explain whether these terms capture distinct learning signals or introduce substantial overlap in optimizing the same structural evidence. Without a more explicit mathematical justification, it is difficult to assess the consistency of the overall objective and determine whether the claimed benefits arise from principled Bayesian inference or from combining potentially overlapping optimization terms. This ambiguity weakens the methodological justification of the proposed framework.

  2. Insufficient Validation of Temporal Regularization under Genuine Community Evolution. The method imposes a KL-divergence penalty between consecutive community-assignment distributions to encourage temporal consistency. While this design can suppress spurious fluctuations, it may also penalize legitimate structural changes, including community splits, merges, births, deaths, and abrupt membership transitions. Although the supplementary material discusses this limitation and reports sensitivity analyses for the temporal regularization coefficient, it does not directly establish that the method can distinguish noise-induced fluctuations from genuine community evolution. Furthermore, the snapshot-wise Hungarian matching used for evaluation does not resolve potential assignment-alignment issues during training, where the temporal regularizer directly compares consecutive assignment distributions. Controlled experiments involving known community transitions are needed to demonstrate that the proposed method preserves genuine structural changes while suppressing transient noise. Without such evidence, the central claim of achieving both temporal stability and dynamic adaptability remains insufficiently substantiated.

Questions for authors
  1. The supplementary material defines a structural training target DtD_t and provides a sparse reconstruction objective in Eqs. (19) and (23). However, the overall objective in Eq. (30) appears to include both the structural reconstruction loss and the variational objective, whose likelihood term is also defined over the observed graph structure. Are these two terms optimizing distinct objectives, or does their combination effectively double-count the same structural evidence? Please clarify their exact mathematical relationship and explain how the likelihood is parameterized and estimated in practice.
  2. The temporal regularizer in Eq. (27) directly minimizes the KL divergence between consecutive community-assignment distributions. Although the supplementary material describes snapshot-wise Hungarian matching for evaluation, this alignment is not used during training. How does the proposed objective ensure that community indices remain semantically aligned across snapshots, given the permutation invariance of community assignments? Furthermore, can the authors provide controlled experiments involving community splits, merges, and abrupt transitions to demonstrate that temporal regularization suppresses spurious fluctuations without suppressing genuine structural evolution?
  3. Please clarify how the number of communities is selected, particularly on real-world networks without ground-truth labels. Additional experiments involving varying community counts, changing community sizes, and node arrivals or departures would help establish the scope of applicability.
2 comments
AuthorsOpenReview ↗

Response to Reviewer wNHe

We appreciate your identification of two issues that influence the interpretation of our results. One pertains to the utilization of structural evidence during the loss. The second inquiry pertains to the preservation of genuine changes by the temporal penalty.

  1. Correlation between the variational likelihood and reconstruction Observations from the identical graph are employed by both phrases. Sparse reconstruction applies binary cross-entropy to the dot product of the assignment probabilities of two nodes in supplementary Eq. (23), after the probabilities have been averaged across posterior draws. A posterior-to-prior KL is combined with an expected negative structural log-likelihood in Supplementary Eq. (19). The structural target is defined in the supplement, and the decoder within that expectation is left as an implicit assumption.

The dot product of the two nodes' membership vectors from posterior draw s would be used as the conditional edge probability in an explicit Bernoulli specification that is consistent with the stated co-membership decoder. The binary cross-entropy of observed edges and sampled non-edges would be added for each draw. The Monte Carlo estimate of the expected negative log-likelihood is obtained by averaging these sums across the S draws.

The two reconstruction terms coincide under this specification for S = 1 and identical sampled pairs. The values obtained by averaging probabilities prior to decoding and averaging losses across draws are generally different when S is greater than 1, despite the fact that the same graph evidence is used. Consequently, Eq. (30) is a composite objective that includes an auxiliary reconstruction loss. The decoder employed in the reported runs, the loss reduction and negative sampling, and the effective weighting of the shared posterior-to-prior KL would be documented. The contribution of posterior inference can be isolated by comparing the full objective with a version that omits the auxiliary reconstruction term and with a matched deterministic head. 2. Training alignment and authentic community transitions The output columns are assigned consistent indices by the shared assignment head. The KL in Eq. (27) promotes continuity in those columns, while semantic correspondence across time is still assumed. Label-dependent evaluation is the sole application of Hungarian matching. Both genuine changes and noise-induced changes are subject to a fixed temporal penalty. The temporal penalty weight was increased from 0.5 to 1.0 in the reported sweep, resulting in a 0.806 to 0.824 increase in CSI and a 0.538 to 0.519 reduction in modularity. This trade-off renders CSI alone insufficient for evaluating transition recovery.

Planted divisions and merges are included in Dynamic LFR and DSBM. The sequences are used to aggregate the scores that are reported. A controlled comparison would involve the use of matched sequences with transient edge corruption or a persistent planted event, such as a split, merge, birth, death, or abrupt membership change. The full model, a version with the temporal penalty set to zero, and a matched deterministic head would be compared. We would assess the detection delay in snapshots, the recovery of changed node-pair memberships around each event, and false assignment changes during stable periods. These measurements are invariant to label renaming due to pairwise co-membership. 3. Handling node turnover and selecting K Within a run, K remains constant. The reference count is implemented when necessary on synthetic benchmarks. The supplementary protocol for real networks specifies label-free selection through temporal stability and validation modularity. For each dataset, we would provide a report that includes the candidate values, the criterion that was employed to select them, and the selected K. The available K columns allow for the modification of group sizes. A common node index and activity markers are employed by the current protocol to identify observed nodes. A sensitivity analysis would involve a separate test for nodes that were not observed during training, as well as a variation in the fitted K, community-size imbalance, and node arrivals and departures.

Thank you for the author's response, but my questions have not been answered, and the supplementary experiments suggested have not been carried out either. Therefore, I have decided to keep my rating unchanged.

Reviewer_aS3yOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Summary: BayesMem-CD distinguishes genuine community evolution from transient noise in dynamic graphs.

Contributions: It combines Bayesian uncertainty estimation, long-term graph memory, and short-term temporal regularization to produce stable, interpretable community assignments.

Validation: Across four synthetic and four real-world datasets, BayesMem-CD achieves the best average community recovery, calibration, and temporal stability while maintaining competitive modularity. Ablations and robustness analyses support the contribution of each component.

Critical strengths

None.

Critical weaknesses
  1. Outdated and poorly contextualized introduction. The introduction cites only work from 2021 or earlier. Although more recent studies are discussed later, key terms and their connection to the paper’s motivation remain unclear. This makes it difficult to assess the method’s novelty and relevance to the current state of the field.

  2. Insufficiently rigorous method presentation. The method section is presented too informally, with important design choices insufficiently explained, making the proposed approach difficult to understand.

Questions for authors
  1. Equation (15) uses p(Yt∣H~t,Θ)p(Y_t \mid \tilde{H}_t,\Theta), but the real-world datasets have no ground-truth communities. What exactly is YtY_t, and how is this likelihood implemented in the unsupervised setting?
  2. The temporal KL regularizer assumes consistent community indices across snapshots, although community labels are permutation-invariant. How does the method align communities over time and handle splits, merges, or label switching?

Nevertheless, I do not expect the authors’ answers to substantially change my overall assessment, as the paper has broader issues that require significant reworking beyond the scope of these clarification questions.

2 comments
AuthorsOpenReview ↗

Response to Reviewer aS3y

Response to Reviewer aS3y

I appreciate your recognition of the implicit assumptions that the introduction and method make. The primary paper should delineate the training observations and provide an explanation of the temporal loss' assumptions regarding community identity.

  1. Introduction and Positioning Recent efforts would be integrated into the motivation. Djurdjevac Conrad et al. (2025) investigate the impact of sparse observations on the detection of dynamic communities, while Yu et al. (2025) investigate the estimation of uncertainty in graph neural networks. A direct precedent for addressing label switching over time is provided by dynamic stochastic block models, such as those developed by Matias and Miele (2017). The problem would be more precisely defined as a result of these comparisons. The observed edges are altered by a transient perturbation, while the underlying partition remains unaltered. The partition that underpins a community transition undergoes a transformation. Our contribution is to assess the recovery, calibration, and stability of the resulting assignments by estimating uncertainty in community assignments from temporally informed graph representations.
  2. Presentation of the Method The method would be presented in computational order, commencing with snapshot encoding and deterministic memory, and proceeding with posterior sampling of the assignment head and the training losses. We would also reconcile the normalized finite-window update specified in the supplementary training procedure with Eq. (7), which presents an exponential recurrence. Before introducing the loss, the revised method would provide a single implementation-level description and define the observed target, the sampled assignment probabilities, and their posterior average.
  3. What is the meaning of Yₜ in Equation (15)? The structural observation set 𝒟ₜ is the likelihood target. It comprises observed edges and sampled non-edge pairs from snapshot t, each of which is represented by a binary indicator. The training targets on real networks are determined by these indicators. The graph is used to infer the K-way community assignments. In the main paper, we would define the pairwise decoder and replace Yₜ with 𝒟ₜ in Eq. (15).
  4. Community alignment, splits, and mergers Throughout a sequence, the same K output columns are supplied by a single assignment head. The temporal KL compares corresponding columns for the same nodes and promotes the consistent use of those columns. The penalty remains unaltered as a result of a shared permutation among all snapshots. Even when the underlying partition is identical, the penalty can be altered by permuting the columns at a single snapshot. The purpose of snapshot-wise Hungarian matching is to map predictions to reference labels for the sole purpose of evaluation. Splits and merges can manifest as modifications to the manner in which nodes occupy the available columns, and sudden modifications result in a KL cost. The aggregate stability is characterized by the present findings. In order to determine whether the method preserves specific transitions, it is necessary to evaluate them in relation to their known onset times.

I thank the authors for their responses. However, they have not provided the relevant experiments or made the corresponding revisions needed to address my concerns. Therefore, I maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.