acceptodds
Under review as a conference paper at ICLR 2027

PHoTE: Photonic Quantum-Walk Encodings for Temporal Link Prediction

Abstract

Jointly learning graph structure and temporal dependencies can make temporal link prediction computationally expensive. Precomputed features avoid structural-encoder optimization, yet their usefulness depends on which information they retain. We introduce PHoTE, a photonic framework combining non-trainable continuous-time quantum-walk encodings with a learned temporal predictor. One- and two-photon statistics at multiple evolution times are projected into identity-sensitive features or reduced to role summaries, yielding four encodings. Features are precomputed per snapshot; only a recurrent readout shared across nodes and a link scorer are trained. We evaluate PHoTE against nine temporal baselines, two classical positional encodings, and two heuristics on nine real networks and 53 synthetic settings from T-GRAB. The two-photon encoding achieves higher mean test average precision than the evaluated temporal baselines and classical encodings on eight of nine real networks. Across its four variants, PHoTE attains or ties the highest mean in five of seven synthetic task–metric comparisons. The identity-sensitive encodings also approach the optimal expected-count reference for stochastic periodicity. After feature precomputation, training takes – seconds per epoch, compared with – seconds for baselines on the same task families. One-photon measurements from a 24-mode processor yield AP comparable to that obtained with analytic features on the full test snapshots of a 13-node Guinea baboon contact network. Results support non-trainable photonic encodings for temporal prediction and demonstrate small-scale hardware feasibility. Although the present encodings are classically simulable, these results motivate exploring richer encodings on future multiphoton quantum processors.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
temporal link prediction, graph representation learning, quantum walks, photonic computing, quantum machine learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_8VHXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper considers the temporal link prediction (TLP) on discrete-time dynamic graphs (DTDGs), a new PHoTE method was proposed. It was claimed to be a photonic framework combing non-trainable continuous-time quantum-walk encodings, which avoids the time-consuming structural encoder optimization. Experiments on real and synthetic dynamic graphs support non-trainable photonic encodings for temporal prediction and demonstrate small-scale hardware feasibility.

Critical strengths

S1. This paper's basic idea about non-trainable photonic encodings for TLP seem interesting.

S2. This paper conducts a great amount of experiments to validate the proposed method's effectiveness.

Critical weaknesses

W1. It was claimed that this paper still adopts the conventional data model of discrete-time dynamic graphs (DTDGs), which may not be considered as a state-of-the-art (SOTA) settings. In contrast, most SOTA techniques for TLP adopt the advanced setting of continuous-time dynamic graphs (CTDGs), with several superiorities beyond DTDGs. Moreover, it was claimed that PHoTE leverages continuous-time quantum walks, which seems to be inconsistent with the setting of DTDGs.


W2. Also as claimed in Section 2, the proposed method can only support the conventional transductive TLP with a fixed node set for all time steps. In contrast, most SOTA TLP approaches (e.g., TGAT, TGN, DyGFormer, etc. in Table 2) can handle the more advanced inductive inference that predicts links for new unseen nodes. Consistent with W1, at least from the perspective of problem statement, this paper cannot be considered as the SOTA.


W3. The main novelty of this paper relies on some photonic preliminaries from recent advances in physics (e.g., "Fock state", "linear-optical circuit", "beam splitters", "phase shifters", "one-photon probabilities", "two-photon coindidence probabilities", "HOM effect", etc.). Although this paper gives definitions for some of these terminologies in Section 2, they were described only in lengthy text, which is hard to read. More importantly, it is also hard to understand how these concepts are linked to TLP on dynamic graphs, hindering a comprehensive evaluation about this paper's novelty. It is suggested to add some toy running examples (e.g., in terms of figures) to further interpret these terminologies.


W4. It was not clearly stated in Section 2 (i.e., problem statement) whether PHoTE need graph attributes inputs (e.g., each node is associated with a feature vector), while some baselines in experiments (e.g., EvoleveGCN, TGAT, TGN, DyGFormer, etc.) relies on such an attribute input, resulting in potential unfair comparison in experiments.


W5. Consistent with W3, there seem no any pseudo-codes to describe the overall pre-processing, training, and inference procedures of PHoTE. As a result, it still hard to understand how does PHoTE exactly work.


W6. While high training efficiency is one of PHoTE's advantages, there seems no formal complexity analysis and comparisons with other methods' complexities to highlight this advantage.


W7. It was claimed that the proposed method can ensure much better training efficiency, while it was also mentioned at the end that "scalability to substantially larger networks remains untested", which may fail to validate this advantage given the close relation between efficiency and scalability.


W8. This paper doses not anonymously provide its code and data to ensure the reproducibility of experiments as some other submissions do.

Questions for authors

See W1-W8.

5 comments

Thank you for the detailed review and for recognizing the motivation and breadth of the experimental study. We agree that the paper should explain its prediction setting, physical construction, computational cost, and reproducibility more clearly. Below we address W1–W8 in order.

W1.

PHoTE studies next-snapshot link prediction on discrete-time dynamic graphs (DTDGs). A continuous-time quantum walk does not change the temporal-data setting. The two time parameters refer to different operations Gt=(V,Et)G_t=(V,E_t) and Ut(τ)=exp⁡(−iτAt)U_t(\tau)=\exp(-i\tau A_t)

Here tt indexes the graph snapshot, while τ\tau is an evolution parameter for a walk on the fixed adjacency matrix AtA_t. We use the walk to construct structural features for each observed snapshot; the temporal readout subsequently processes those features across snapshots. Given G1,…,GtG_1,\ldots,G_t, the prediction target is the edge set of Gt+1G_{t+1}.

DTDG and CTDG are different representations of temporal data, not a hierarchy in which one makes the other invalid. UTG and T-GRAB evaluate both snapshot-based and event-based temporal methods. We have clarified this distinction in Section 2 and made the prediction setting explicit in the Introduction.

W2.

We clarify that PHoTE is designed for transductive temporal link prediction on a fixed node set. Generalization to unseen nodes is not the research question addressed in this work. Our focus is on whether fixed quantum-walk representations can capture temporal graph information and support accurate next-snapshot link prediction.

The DTDG/CTDG distinction concerns the representation of temporal change. However, transductive/inductive refer to different settings of generalization beyond training nodes. Methods such as TGAT, TGN, and DyGFormer can have inductive capabilities, but in our comparison they are evaluated on the same specified transductive target as PHoTE.

W3/W5.

We thank the reviewer for the suggestions on improving the presentation. We have clarified the methodological workflow through Figure 1 and Algorithm 1, covering feature precomputation, chronological training, and test-time prediction.

The photonic concepts describe how the graph-dependent statistics can be physically obtained. Our methodological contribution lies in constructing node representations from one-photon transition probabilities and edge-aggregated two-photon coincidence statistics, and in studying how identity-sensitive profiles and role summaries support temporal prediction.

To make these representations more intuitive, Appendix F.2 uses simple graph structures to illustrate the information captured at different evolution times. Appendix F.3 (Figure 8 and Table 36) provides a concrete example in which node 17 changes its interaction partner from node 42 to node 88. The identity-sensitive features change their destination coordinates, whereas the role summaries remain unchanged.

We have also added control experiments in Appendix D.1, comparing the quantum-walk encodings with node identifiers, adjacency vectors, heat kernels, and personalized PageRank under the same temporal readout. These comparisons help distinguish the contribution of quantum-walk features from that of node identity and simpler structural information.

Together, these revisions clarify how the physical construction produces features used for temporal link prediction and how the proposed representations differ from existing alternatives.

W4.

PHoTE requires only binary snapshot adjacency matrices and does not use observed node or edge attributes. In the main experiments, the temporal GNN baselines receive one-hot node identifiers following the T-GRAB protocol. No observed node or edge attributes are supplied in these comparisons. The methods are evaluated on the same prediction target and chronological splits, although their input representations and architectures differ.

LapPE and RWPE provide an additional controlled comparison: like PHoTE, they construct structural features from adjacency matrices and use the same temporal readout.

For the additional TGB experiments in Appendix A.5, the continuous-time baselines also receive the benchmark edge attributes, whereas PHoTE continues to use only adjacency matrices. We have made this distinction explicit to clarify the information available to each method.

W6.

We have added a computational-cost analysis in Appendix G.1, separating one-time feature construction from the training and inference costs of the temporal readout.

For MM nodes, mt=∣Et∣m_t=|E_t| edges in snapshot tt, LL evolution times, and projection dimension DD, the direct dense per-snapshot encoding costs are

T1ph(t)=O(LM3+LM2D),T_{\mathrm{1ph}}(t)=O\left(LM^3+LM^2D\right), T2ph(t)=O(LM3+LmtM2+LM2D).T_{\mathrm{2ph}}(t)=O\left(LM^3+Lm_tM^2+LM^2D\right).

These account for matrix exponentials, explicit accumulation of the two-photon coincidence table, and projection. They are upper bounds for the direct implementation, not universal lower bounds.

Let dd be the encoded node-feature dimension, hh the common LSTM/MLP hidden width, BB the number of candidate pairs scored in one pass, and TT the number of snapshots. A representative readout cost for one non-overlapping pass is

O ⁣(TM(dh+h2)+Bh2).O\!\left(TM(dh+h^2)+Bh^2\right).

Here d=Dd=D for projected profiles and d=L(Ktop+4)d=L(K_{\mathrm{top}}+4) for role summaries. Reprocessing overlapping temporal windows adds corresponding readout work. The fixed graph encodings are constructed once and reused over training epochs, whereas a learned structural encoder generally incurs training-time forward and backward computation. LapPE and RWPE paired with the same readout share the same readout complexity when dimensions match.

The key training-efficiency advantage comes from computing structural features once and reusing them across epochs, rather than repeatedly differentiating through a structural encoder. This distinction also applies to LapPE and RWPE, which share the same readout complexity when feature dimensions match. The quantum-walk encodings are therefore not claimed to provide an intrinsic acceleration of the training stage.

For empirical comparison with the heterogeneous temporal GNN architectures, Appendix G.2, Tables 37–40 report per-epoch training times and separate one-time preparation times. All methods were timed on NVIDIA H100 GPUs with PyTorch 2.9. The per-epoch measurements exclude feature preparation, validation, and testing. We do not equate lower per-epoch training time with lower total computational cost in every setting.

W7.

We appreciate the reviewer's suggestion to evaluate PHoTE on larger graphs. We have extended the evaluation to two TGB datasets, tgbl-wiki (9,227 nodes) and tgbl-flight (18,143 nodes), using the official TGB splits, negative samples, and mean reciprocal rank (MRR) metric. The results are reported in Appendix A.5, Table 7.

On tgbl-flight, the two-photon and one-photon encodings achieve the two highest mean MRR scores among the methods reported in Table 7, reaching 0.861 ± 0.001 and 0.852 ± 0.000, respectively. The strongest evaluated temporal GNN, GC-LSTM, achieves 0.842 ± 0.002.

On tgbl-wiki, one-photon achieves 0.512 ± 0.004, outperforming every evaluated discrete-time GNN. However, EdgeBank (0.569), CTAN (0.605), and DyGFormer (0.735) achieve higher scores. Thus, PHoTE remains competitive at larger scales, although its performance varies across datasets.

We also report feature-construction costs. On one GPU, precomputing features for the full tgbl-flight sequence takes approximately 26 minutes for one-photon and 90 minutes for two-photon. For tgbl-wiki, isolated nodes can be handled analytically, reducing the matrix-exponential computation to approximately 150 active nodes per hourly snapshot.

These results extend the evaluation substantially beyond the original networks, while also illustrating the cost of feature precomputation. They do not establish an end-to-end runtime advantage in every setting or demonstrate large-scale photonic hardware execution.

We have accordingly updated the Limitations section (Section 7) to reflect the larger datasets while retaining the limitation that scalability to substantially larger networks and photonic devices remains untested.

W8.

We thank the reviewer for emphasizing reproducibility. We have made an anonymized replication package available to reviewers:

Replication package: https://anonymous.4open.science/r/photonic_quantum_walk-32E8

We have also added Algorithm 1, which describes the complete PHoTE pipeline, and expanded the implementation details in Appendix G. Together, these resources are intended to facilitate independent verification and reproduction of our experimental results.

Thank you again for these comments. We hope the clarifications and additional evidence address your concerns. If they do, we would be grateful if you would reconsider your assessment.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

We hope to have addressed your concerns and answered your questions. We remain available if you have any more questions during the end of the discussion period. If we properly addressed your concerns and questions, we kindly ask you to consider raising your score.

Reviewer_cACfOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
1
Soundness
3(+1)
Read review
Paper summary

The paper proposes a novel method (PHoTE) for link prediction in temporal graphs. Using continuous-time quantum walks, PHoTE encodes each snapshot. The authors state that the method is reproducible and can be simulated with a linear-optical processor with one mode that implements the quantum walk. The proposed features are built from one- or two-photon output statistics, either as identity-sensitive profiles or as permutation-invariant role summaries. The authors highlight the method's efficiency: only a shared LSTM and an MLP are trained. Experiments are conducted against several baselines and on several datasets for link prediction in temporal graphs.

Critical strengths
  1. Using quantum walks as fixed encodings for temporal link prediction is novel. Quantum-walk features have been used in related work (e.g., static graph similarity), but, to the best of my knowledge, not for discrete-time temporal link prediction. The two-photon coincidence features are an original construction.
  2. The synthetic analysis provides reference points that make the results interpretable. The paper is also clear that the proposed encodings can be reproduced with classical simulations.
  3. Evaluation is broad and includes learned baselines as well as classical encodings. The hardware experiment provides a nice demonstration.
Critical weaknesses
  1. I think the experiments do not isolate the gain arising from PHoTE's quantum-walk dynamics. In the manuscript, PHoTE encodings appear to preserve a correspondence between feature values and the corresponding destination nodes. In contrast, RWPE does not carry this information, and LapPE uses spectral coordinates that may change across snapshots. Consequently, I think that the reported comparison does not show how much of the improvement comes from the quantum walk itself and how much comes from the information retained in the representation.
  2. The authors say that continuous-time models process timestamped events, but the real datasets and targets are presented as (aggregated) snapshots. The authors detail the aggregation procedure in the appendix, but it remains unclear whether the baselines receive the original event stream or events reconstructed from the snapshots. Moreover, on T-GRAB, several entries in Tables 10 and 11 are close to the F1 score obtained by predicting every candidate as an edge. This raises the question of whether the low scores reflect poor ranking performance or predictions that are poorly calibrated for the fixed threshold. Reporting a threshold-free metric and the proportion of predicted positives would help clarify this. Finally, the PHoTE variants appear to have very different strengths across task families. I think this is an interesting finding, but it remains unclear how reliably an encoding can be selected based on the validation data. The authors should specify the model selection procedure used.
  3. Presentation. The manuscript is really hard to follow: sentences are dense and heavily hedged. Much of the evidence supporting the main claims is relegated to the appendix, whose writing suffers from the same issues. The writing would need substantial revision.
Questions for authors
  1. Can the authors address W1 and present a solid ablation that isolates the contribution of the information retained by the PHoTE encodings? Can the authors formally prove that this additional information does not influence the predictions, or empirically assess whether it influences the models?
  2. Can the authors provide additional metrics beyond F1 to help assess the quality of the link predictions?
  3. Can the authors detail the validation procedure adopted to select the encodings?
  4. Can the authors detail precisely how continuous-time baselines were compared on snapshot-based events, beyond what is stated in Appendix A?
5 comments

We thank the reviewer for the detailed assessment and for recognizing the novelty of applying fixed quantum-walk features to discrete-time temporal link prediction, the original two-photon construction, and the breadth of our evaluation. Your constructive feedback has helped us strengthen the experimental evidence, better distinguish the contributions of quantum-walk dynamics and destination-node correspondence, and clarify our evaluation protocol and presentation. We have conducted additional experiments and revised the manuscript to address these points.

W1/Q1.

We thank the reviewer for highlighting the importance of distinguishing quantum-walk dynamics from retained destination-node correspondence. Our goal is to assess its contribution separately from the predictive benefits of quantum-walk propagation, rather than claim that correspondence has no effect.

We added five control encodings in Appendix D.1. Setting all evolution times to zero gives Ut(0)=IU_t(0)=I and Et(2)(0)=AtE_t^{(2)}(0)=A_t. The resulting features are projected node identifiers for the one-photon encoding and projected adjacency vectors for the two-photon encoding. We also evaluate learnable node embeddings and two classical diffusion controls, the heat kernel and personalized PageRank (PPR). These controls use the same LSTM-based temporal readout and evaluation protocol. The learnable embedding control additionally trains its node embeddings.

Table 1. Selected results across 53 synthetic and 11 real-data settings (three seeds). Synthetic results report mean test F1F_1; real networks report mean AP.

EncodingMem.Sto. p=0.5p=0.5Long-rangeCause-effectReal AP
One-photon0.9880.6570.5530.9100.287
Two-photon0.9460.6570.6790.9600.321
Q1-role0.9100.3220.9220.9760.254
Q2-role0.9090.3740.9070.9780.261
Projected identifier0.2110.3530.0590.2720.242
Learnable embedding0.2000.4320.0590.2750.257
Projected adjacency0.9500.5960.5850.9040.313
Heat kernel0.9880.6570.5070.8250.253
PPR0.9760.6570.3920.8660.263

From this results, we show that both fixed and learnable identity embeddings perform substantially worse than PHoTE on synthetic temporal reasoning tasks. Node identity alone is insufficient. Both fixed and learnable identity embeddings perform substantially worse than PHoTE on synthetic temporal reasoning tasks. For example, on memorization, the projected identifier achieves 0.211 compared with 0.988 for one-photon.

This results show that quantum-walk features can improve prediction beyond the tested adjacency and diffusion controls. Compared with the zero-time adjacency control, two-photon features improve stochastic periodicity (0.596 to 0.657), long-range memory (0.585 to 0.679), and cause-and-effect (0.904 to 0.960). The improvement is smaller on real networks (0.313 to 0.321). Classical diffusion matches PHoTE on stochastic periodicity, indicating that the benefits of quantum-walk dynamics are task-dependent.

The results imply that destination node correspondence is not always beneficial for strong prediction. Q1-role and Q2-role remove the correspondence between non-self feature coordinates and destination nodes, yet outperform their identity-sensitive counterparts on long-range memory and cause-and-effect. Conversely, identity-sensitive encodings perform better on memorization, counting, and stochastic periodicity.

The complete control results and standard deviations are included in Appendix D.1, Tables 16–23. We will also highlight the main findings in Section 6.

W2/Q2.

We thank the reviewer for suggesting threshold-free metrics to better distinguish poor ranking from threshold-related effects. We use F1 for consistency with T-GRAB and have added average precision (AP), a threshold-free ranking metric, for the synthetic benchmarks in Appendix B.3. We also measured the fraction of candidate pairs predicted as positive at the original threshold of 0.5.

Table 2. Selected long-range memory results: family-mean test F1, AP, and fraction of candidates predicted as positive at the original threshold.

MethodF1APPredicted positives
GCN0.2380.25652.1%
GAT0.2670.28554.7%
TGAT0.7430.79313.0%
TGN0.7320.77610.6%
DyGFormer0.7570.80911.8%
Q1-role0.9220.9019.1%
Q2-role0.9070.9138.0%

Only 3% of candidate edges are positive on long-range memory. The low F1 scores of GCN and GAT therefore reflect poor ranking as well as excessive positive predictions, rather than merely an unsuitable decision threshold. In contrast, the role summaries achieve the highest AP among the evaluated methods, supporting their effectiveness independently of the F1 threshold.

On cause-and-effect, Q1-role reaches AP 0.990, close to TGN (0.999) and TGAT (0.998). These results confirm that the role summaries remain competitive under threshold-free evaluation. On cause-and-effect, Q1-role achieves AP 0.990, close to TGN (0.999) and TGAT (0.998). The relative ordering of the two role summaries also differs between F1 and AP, reinforcing the value of reporting both metrics.

The AP results are in Appendix B.3, Tables 14–15. We will add the predicted-positive fractions to the supplementary results.

W2/Q3. We distinguish two choices between checkpoint and encoding section.

Checkpoint selection. The default readout training budget is 600 epochs, with validation every ten epochs. The selected checkpoint maximizes the relevant validation metric: F1F_1 on synthetic tasks and AP on the real-network AP protocol. Validation metrics are computed per snapshot and averaged equally over the validation split. For counting and stochastic periodicity the relevant benchmark candidate set is used; for long-range and cause-and-effect, candidate pairs involve the memory node. Tied validation scores retain the earlier checkpoint.

Encoding selection. The main experimental tables report all four PHoTE encodings independently. They therefore do not rely on selecting the best encoding using test performance.

To evaluate whether validation can reliably select an encoding, we conducted an additional analysis over 64 settings (53 synthetic and 11 real-data settings), each with three seeds. For each setting and seed, we select the encoding with the highest validation score, breaking ties in the order one-photon, two-photon, Q1-role, and Q2-role. This produces 192 selections.

Validation-selection outcomeCount
Selected encoding also best on test163 / 192 (85%)
Selected encoding within 0.01 of test-best188 / 192 (98%)
All three seeds select the same encoding35 / 64

On counting and stochastic periodicity, validation selects one of the identity-sensitive encodings; on cause-and-effect, it selects a role summary. The results show that validation usually identifies a strong encoding, while exact encoder choice is not stable across every seed.

W2/Q4.

We agree that the distinction between original timestamped events and snapshot-derived events should be explicit.

Our primary evaluation targets next-snapshot link prediction. For the nine main real interaction networks, we first aggregate the original event records into undirected, unweighted snapshots using the time windows described in Appendix A.1. The synthetic benchmarks are already generated as snapshot sequences.

Following the official T-GRAB implementation, the continuous-time baselines receive an event stream constructed from these snapshots, not the original fine-grained event stream. Each undirected edge {u,v} in snapshot t becomes two directed events, (u,v,t) and (v,u,t), both timestamped by the snapshot index.

For every target snapshot, all candidate pairs are scored before any events from that snapshot are added to the model's history or used to update its state. Predictions therefore use only preceding snapshots. All methods follow the same chronological splits and snapshot-level prediction targets.

This protocol enables a comparison on the same snapshot-based task. It also removes intra-window timing information that continuous-time models could use in their native event-level settings. We do not interpret these results as a comparison of the models on their original event-stream prediction tasks.

We have clarified this conversion in Section 5 and the real-data aggregation procedure in Appendix A.1. The additional tgbl-wiki and tgbl-flight experiments under the TGB protocol are reported separately in Appendix A.5.

W3.

We thank the reviewer for the suggestions on improving the presentation.

We would like to emphasize that PHoTE is a graph representation learning method that can be understood through its mathematical formulation, without requiring prior knowledge of photonic computing. The photonic background provides the physical interpretation and implementation of the proposed encodings, while the methodological contribution lies in constructing and evaluating these representations for temporal link prediction.

We have made several changes in the revised manuscript.

First, Algorithm 1 in Section 4 now presents the complete pipeline, including feature precomputation, chronological LSTM training, checkpoint selection, and test inference. This makes the computational procedure explicit without requiring the reader to follow the photonic interpretation first.

Second, Appendix F.2 provides concrete examples showing what the quantum-walk probabilities capture at different evolution times. These connect the mathematical definition to node degree, surrounding graph structure, and longer-time propagation.

Third, we added a worked example in Appendix F.3 and Figure 8 to explain the distinction between identity-sensitive profiles and role summaries. In the example, node 17 interacts with node 42 in one snapshot and node 88 in another. The original photon profile places the corresponding probability at different destination coordinates, whereas the role summary retains the same structural description. This illustrates why role summaries may help recognize recurring structural patterns when the participating nodes change.

We also clarify that each graph node maps to one optical mode, so a graph with M nodes is represented by an M-mode linear-optical transformation.

To make the central evidence more accessible, we will add a concise version of the worked example to Section 4 and a summary of the new control experiments to Section 6. These changes will reduce the need to consult the appendix to understand the method and its main findings.

Thank you for highlighting these issues. We hope the new controls, ranking-based evaluation, and more explicit model-selection analysis address your concerns. If you find that these additions satisfactorily resolve your concerns, we would be grateful if you could reconsider your rating.

Comments on the paper

5 comments

We thank all three reviewers for their careful reading and constructive feedback. Based on their suggestions, we have substantially strengthened the manuscript through additional experiments, more detailed analysis, and clearer presentation. All changes in the revised manuscript are highlighted in blue for ease of review. We summarize the main updates below.

1. Additional baselines and control experiments. We added five control encodings based on node identifiers, learnable embeddings, adjacency vectors, heat kernels, and personalized PageRank (cGfL: W1/Q1; cACf: W1/Q1; 8VHX: W3/W5). These experiments help identify the contributions of quantum-walk propagation and node identity. We also added GraphMixer as a baseline across all synthetic and real-data settings (cGfL: W2). The new comparisons provide further evidence for the effectiveness of PHoTE.

2. Extended evaluation and analysis. We added average precision (AP) for synthetic benchmarks and analyzed validation-based encoding selection (cACf: W2/Q2–Q3). We also conducted ablation studies on evolution times, projection dimension, and role-summary size (cGfL: W4). These results provide a more complete understanding of PHoTE's performance and parameter choices.

3. Computational efficiency and scalability. We added computational complexity analysis and expanded runtime measurements to include both feature preparation and training (cGfL: W3/Q3; 8VHX: W6). We further evaluated PHoTE on two larger TGB datasets, tgbl-wiki (9,227 nodes) and tgbl-flight (18,143 nodes), where it achieves competitive performance (cGfL: W3/Q3; 8VHX: W7).

4. Methodology and presentation. We clarified the prediction setting, evaluation protocol, and baseline inputs (8VHX: W1/W2/W4; cACf: W2/Q4). We also added Algorithm 1, improved the pipeline description, and included worked examples to explain the quantum-walk features and their use in temporal prediction (cACf: W3; 8VHX: W3/W5). In addition, we extended the hardware study to full-test-sequence two-photon measurements, obtaining results comparable to analytic features.

5. Reproducibility. We expanded the implementation details and released an anonymized replication package (8VHX: W8):

https://anonymous.4open.science/r/photonic_quantum_walk-32E8

We believe these additions strengthen the experimental support for PHoTE and make its methodology and contributions clearer. We thank the reviewers again for their valuable suggestions.

We thank the reviewer for the detailed assessment and for recognizing the clarity of our methodology, the transparency of our experimental reporting, and the completeness of our hardware evaluation. Your comments have helped us strengthen the experimental evidence, distinguish quantum-walk dynamics from destination-node correspondence, and clarify the computational costs and limitations of PHoTE.

W1/Q1.

We thank the reviewer for highlighting the importance of distinguishing quantum-walk dynamics from retained destination-node correspondence.

We compare encodings that preserve destination-node correspondence while removing quantum-walk propagation. At zero evolution time, Ut(0)=IU_t(0)=I and Et(2)(0)=AtE_t^{(2)}(0)=A_t. Thus, one-photon features reduce to projected node identifiers, while two-photon features reduce to projected adjacency vectors. We further include learnable node embeddings and two classical diffusion baselines: heat kernels and personalized PageRank (PPR; Jeh and Widom, 2003). All controls use the same LSTM-based temporal readout and evaluation protocol.

Table 1. Selected results across 53 synthetic and 11 real-data settings (three seeds). Synthetic results report mean test F1F_1; real networks report mean AP.

EncodingMem.Sto. p=0.5p=0.5Long-rangeCause-effectReal AP
One-photon0.9880.6570.5530.9100.287
Two-photon0.9460.6570.6790.9600.321
Q1-role0.9100.3220.9220.9760.254
Q2-role0.9090.3740.9070.9780.261
Projected identifier0.2110.3530.0590.2720.242
Learnable embedding0.2000.4320.0590.2750.257
Projected adjacency0.9500.5960.5850.9040.313
Heat kernel0.9880.6570.5070.8250.253
PPR0.9760.6570.3920.8660.263

From these results, we show that both fixed and learnable identity embeddings perform substantially worse than PHoTE on synthetic temporal reasoning tasks.

These results show that quantum walk propagation adds predictive value beyond adjacency. Compared with the zero-time adjacency control, two-photon features improve stochastic periodicity (0.596 to 0.657), long-range memory (0.585 to 0.679), and cause-and-effect (0.904 to 0.960). The improvement is smaller on real networks (0.313 to 0.321). Classical diffusion matches PHoTE on stochastic periodicity, indicating that the benefits of quantum-walk dynamics are task-dependent.

The results imply that destination node correspondence is not always beneficial. Q1-role and Q2-role remove the correspondence between non-self feature coordinates and destination nodes, yet outperform their identity-sensitive counterparts on long-range memory and cause-and-effect. Conversely, identity-sensitive encodings perform better on memorization, counting, and stochastic periodicity.

All these experiments show that both graph propagation and representation design matter with task-dependent contributions. We have added the control definitions, family-level comparisons, and detailed per-setting results in Appendix D.1 (Tables 16–23), including the variability across three seeds.

W2.

Thank you for identifying GraphMixer as relevant prior work and a missing baseline. GraphMixer employs temporal-link processing with a trained MLP-Mixer and neighbor aggregation. PHoTE studies fixed, graph-dependent one- and two-photon statistics, including their organization into identity-sensitive profiles or role summaries. We have completed the GraphMixer comparison across all 53 synthetic settings and 11 real-data settings, with three seeds per configuration. We use the public DyGLib implementation within the same T-GRAB evaluation pipeline, following the data splits and input protocol used for the other continuous-time baselines.

Table 2. GraphMixer comparison across the synthetic task families and real networks. Results are averaged over three seeds.

MethodMem.Count.Sto. p=0.9p=0.9Sto. p=0.5p=0.5Long-rangeCause-effectReal AP
GraphMixer0.8970.8420.5260.3390.8180.9290.243
Best PHoTE encoding per family0.9880.9370.9370.6570.9220.9780.321

For the completed comparisons, the best PHoTE family mean exceeds GraphMixer. In the measured synthetic runs, GraphMixer requires at least approximately 3.5 hours per run to reach its retrospectively selected checkpoint on average within each task family, compared with at most approximately 11 minutes, including preparation, for the PHoTE encodings.

GraphMixer is now included in the related-work discussion (Section 3), the main benchmark comparisons (Section 6.2), and the detailed results and runtime tables (Appendices A, B, and G).

W3/Q3.

The submitted version separately reported training time per epoch and one-time preparation time. For instance, two-photon preparation on tgbn-genre required 628.739 seconds. That preparation includes startup, imports, data loading, encoding construction, and model initialization; the per-epoch figures excluded preparation, validation, and testing.

We have now measured training time up to the checkpoint selected retrospectively by validation, using one dedicated NVIDIA H100 GPU for each method. We report the one-time preparation cost separately and alongside the sum of preparation and training-to-checkpoint time. For tgbn-genre, the two-photon pipeline requires approximately 631 seconds of preparation plus 136 seconds of training, or 767 seconds (under 13 minutes) under this timing definition. The corresponding temporal GNN measurements range from approximately 26 minutes (GC-LSTM) to 19 hours (TGCN). On the synthetic task families, each PHoTE encoding reaches its selected checkpoint faster than the fastest temporal GNN, including preparation, in five of six families. The exception is stochastic periodicity with p=0.9p=0.9, where EvolveGCN reaches a selected checkpoint earlier but has ROC-AUC 0.500.

We have added the complexity of PHoTE. Let MM be the node count, mt=∣Et∣m_t=|E_t| the number of snapshot edges, LL the number of evolution times, and DD the output projection dimension. The direct dense per-snapshot costs are

T1ph(t)=O(LM3+LM2D),T_{\mathrm{1ph}}(t)=O\left(LM^3+LM^2D\right),

T2ph(t)=(LM3+LmtM2+LM2D).T_{\mathrm{2ph}}(t)=\left(LM^3+Lm_tM^2+LM^2D\right).

The terms account for dense matrix exponentials, the explicit edge-aggregated two-photon coincidence table, and projection of the resulting node features. The encodings are computed once and reused across training epochs.

We additionally evaluated two larger TGB benchmarks tgbl-wiki (9,227 nodes) and tgbl-flight (18,143 nodes) using the official TGB MRR protocol (https://tgb.complexdatalab.com/). The latter has approximately 12 times as many nodes as the largest primary network in the original submission. All trained methods were evaluated over five seeds. The temporal GNN baselines use their published configurations, with selected hyperparameters tuned on validation MRR. For PHoTE, the evolution times and readout hyperparameters were selected by validation MRR in these additional experiments.

Table 3. Test MRR, mean ±\pm standard deviation, under the TGB protocol.

Methodtgbl-wikitgbl-flight
LapPE0.438 ± 0.0020.312 ± 0.065
RWPE0.316 ± 0.0050.797 ± 0.002
One-photon0.512 ± 0.0040.852 ± 0.000
Two-photon0.381 ± 0.0120.861 ± 0.001
Q1-role0.427 ± 0.0020.819 ± 0.004
Q2-role0.423 ± 0.0040.830 ± 0.002

PHoTE achieves competitive performance on both larger TGB benchmarks. On tgbl-flight, two-photon and one-photon achieve the two highest test MRRs among all reported methods, at 0.861 and 0.852, respectively. On tgbl-wiki, one-photon achieves a test MRR of 0.512, ranking 3rd among the evaluated trained methods, behind DyGFormer (0.735) and CTAN (0.605). These results extend our evaluation to larger networks and show that the performance advantage is task-dependent. Constructing features for the full tgbl-flight sequence took approximately 26 minutes for one-photon/Q1-role and 90 minutes for two-photon/Q2-role on one GPU.

We have added a computational complexity analysis in Appendix G.1, distinguishing the one-time encoding cost from the cost of training the temporal readout.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.