acceptodds
Under review as a conference paper at ICLR 2027

AutoFed: Personalized Federated Traffic Prediction via Adaptive Prompt

Abstract

Accurate traffic prediction is essential for Intelligent Transportation Systems, including ride-hailing, urban road planning, and vehicle fleet management. However, due to significant privacy concerns, most existing methods rely on local training, resulting in data silos and limited knowledge sharing. Federated Learning (FL) enables privacy-preserving collaborative training, but standard FL struggles with non-independent and identically distributed (non-IID) data across clients. Personalized Federated Learning (PFL) has emerged as a promising paradigm, yet current PFL frameworks require extensive adaptation for traffic prediction tasks, such as graph feature engineering and network architecture design. Moreover, many prior methods depend on dataset-specific hyper-parameter optimization, which impedes practical deployment. To address these challenges, we propose AutoFed, a novel PFL framework for traffic prediction that significantly reduces reliance on manual hyper-parameter tuning and feature design. Inspired by prompt learning, AutoFed introduces a federated representor that employs a client-aligned adapter to distill local data into a compact, globally shared prompt matrix. This prompt then conditions a personalized predictor, allowing each client to benefit from cross-client knowledge while maintaining local specificity. Extensive experiments on real-world datasets demonstrate that AutoFed consistently achieves superior performance across diverse scenarios. The code of this paper is provided at anonymous repository https://anonymous.4open.science/r/AutoFed-3440 .

Primary area:
learning on time series and dynamical systems
Keywords:
Traffic Prediction, Personalized Federated Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5DW5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work introduce AutoFed, which is a personalized federated learning framework for traffic prediction to handle clients with non-uniform data distributions and varying graph sizes.

The main contribution of this paper is an input-dependent prompt mechanism. The mechanism shares only feature-level auto-encoder and adapter parameters while keeping graph-dependent components and predictors local. It aim at making per-client communication independent of graph size. Experimental results claim competitive accuracy and reduced communication compared to baselines.

Critical strengths

Two main critical strengths.

The first one is an effective personalized sharing mechanism (AutoFed). It uses input-dependent prompts to connect shared modules with private predictors. It enables cross-client collaboration while preserving adaptation to local traffic patterns.

The second one is it provides support for heterogeneous graph sizes with low communication overhead. The shared parameters are independent of the number of nodes. It allows different-sized clients can collaborate.

Critical weaknesses

The benefit of federated sharing should be clearly isolated. It is suspected that the improvements may come from the additional local prompt-generation pathway rather than the cross-client collaboration. In the experiments, it would be better to include a complete AutoFed model trained without sharing (under the same training budget); this would help distinguish these effects.

Also, the baseline fairness needs clarification. It feels like the fixed configurations do not necessarily suit all methods equally. The paper should also explain hyperparameter selection, validation budgets, and the impact of modifying training procedure.

Questions for authors
  1. Is that possible to report a no-sharing control using the complete AutoFed architecture (including the representor, adapter, and reconstruction objective) under the same training budget?

  2. What are the exact input-feature and latent dimensions of the AE for each task?

  3. Please clarify some important setting involved in the experimental protocol; for example the history/prediction lengths, client node counts, PEMS partitioning, aggregation weights..

4 comments
AuthorsOpenReview ↗

Response Part 1

Thank you for your careful review and constructive comments. We have added experiments and clarified the method, analysis, and experimental protocol. Our point-by-point responses follow.

W1 / Q1: Complete-model local control under the same training budget.

Thank you for your insightful comment. We added the complete AutoFed model trained locally without sharing (w/o FL) to Section 4.2, Tables 1–2. Its performance is similar to the original local AGCRN baseline, whereas federation lowers MAE in every reported setting. This supports cross-client learning as the main source of the improvement.

Here we show one setting from each task, S3 and PEMS08; the complete results are in Table 1 for TDP and Table 2 for TFP. Local AGCRN is the original baseline, and w/o FL retains the complete AutoFed architecture and objective. Entries follow the mean ± standard deviation convention in those tables.

SettingMethodMAERMSEMSEMAPE (%)
S3 UberLocal AGCRN3.45±0.055.43±0.1033.69±1.5639.45±0.13
S3 Uberw/o FL3.44±0.035.42±0.0833.42±0.8139.32±0.20
S3 UberAutoFed3.36±0.045.27±0.0731.56±0.8938.82±0.49
S3 LyftLocal AGCRN2.00±0.043.06±0.069.61±0.3949.16±0.19
S3 Lyftw/o FL1.94±0.003.01±0.019.38±0.0750.21±0.30
S3 LyftAutoFed1.85±0.022.82±0.048.40±0.1950.33±0.29
PEMS08Local AGCRN18.65±0.3627.64±0.53790.42±23.9012.44±0.45
PEMS08w/o FL18.97±0.2127.73±0.20793.21±17.5413.33±0.33
PEMS08AutoFed16.83±0.1725.62±0.21684.63±14.1611.10±0.26

The local prompt pathway transforms the same observed history and provides little additional forecasting benefit on its own. The gain emerges when its mappings learn from other clients through federation. Appendix A.3 defines the control; the sharing analysis in Section 4.2 is marked in blue.

W2: Baseline settings, selection budget, and training procedures.

Thank you for your insightful comment. We used each baseline's best-performing configuration reported in its paper, following the official implementation for method-specific settings. No method, including AutoFed, received additional hyperparameter tuning in our experiments. We removed the optional curriculum-learning schedule from the FedTPS implementation to compare the core methods without additional training enhancements.

Curriculum learning is an optional, general-purpose training heuristic included in the official FedTPS repository, rather than a component of the traffic-pattern-sharing method proposed in the paper. Such heuristics can also be applied to AutoFed and other compatible models to improve training. Combining multiple enhancements makes it difficult to distinguish gains from the proposed method from gains due to its training recipe. We therefore use a clean comparison without adding curriculum learning or stacking extra performance-enhancement heuristics, while retaining each method's core algorithm and method-specific hyperparameters.

Section 4.1 and Appendix A.2 specify the configuration sources and the curriculum-learning choice. Every method reports the test result from its best-validation round, and results are mean ± standard deviation over three runs. The common checkpoint-selection rule is marked in blue in Appendix A.2.

Q2: Exact AE input and latent dimensions.

Thank you for your insightful comment. We added the exact AE dimensions to Appendix A.1, Table 4, and clarified the whole-window mapping in Section 3.4. The AE is 108 → 768 → 108 for TDP and 12 → 768 → 12 for PEMS. It expands each node's observed history before reconstruction in both tasks.

TaskHistory TInput features KHidden width kFlattened input → latent → reconstruction
TDP12964108 → 768 → 108
PEMS1216412 → 768 → 12

The map is T × K → T × k → T × K; the latent is reshaped into a 12-step sequence with 64 features per node for graph encoding. Its adapter output is then concatenated with the predictor encoder state to initialize the decoder. Reconstruction learns a temporal-feature representation that supplies this additional forecasting condition. The whole-window explanation in Section 3.4 and the initialization in Section 3.2 are marked in blue.

AuthorsOpenReview ↗

Response Part 2

Q3: History, partitions, client sizes, and aggregation.

Thank you for your insightful comment. We consolidated these settings in Appendix A.1, including task dimensions in Table 4, exact PEMS client node counts in Table 5, and split sizes in Table 6. Both tasks use 12-step histories and forecasts with stride one; PEMS uses fixed FedTPS METIS partitions, with full participation and node-count-weighted aggregation. Appendix B gives the algorithm.

TDP S0 has one city-wide client per operator; S1 has five Uber borough clients, S2 five Lyft clients, and S3 all ten regional clients. Per operator, the Bronx, Brooklyn, Manhattan, Queens, and Staten Island contain 43, 61, 69, 67, and 20 nodes. S0 combines these into 260 nodes per client.

PEMS03/04/08 have 358/307/170 sensors, respectively. The FedTPS loader partitions each dataset into 2, 4, 6, 8, or 10 disjoint METIS parts and caches the assignment for reuse across runs. Table 5 lists every partition's client sizes; the fixed-partition protocol is marked in blue. Appendix A.1 also gives the time splits, preprocessing, and evaluation convention.

Shared updates use ni/∑jnjn_i/\sum_j n_j, giving larger local graphs proportionally more weight. Graph modules, predictors, and BN remain private. Evaluation averages clients equally, measuring average client performance separately from the node-weighted training update.

Thanks for the clarification and additional experiments. My concerns have been partially addressed. However, while I appreciate the methodology design, I remain uncertain with the technical contribution beyond existing Fed-sharing framework is sufficient to meet the expectation of this venue. Therefore, I will maintain my current rating. Thanks for your understanding.

Thank you for your further comments. We fully understand your final decision and appreciate your suggestions for improving our manuscript.

We would like to clarify the contribution beyond generic federated parameter sharing. AutoFed does not claim novelty from AGCRN, auto-encoders, FedBN, or prompt conditioning in isolation. The technical contribution is a traffic-specific sharing interface for heterogeneous graph predictors: each client uses its private adaptive graph encoder to transform the current node and history window into an input-dependent prompt, while only the node-count-independent auto-encoder and adapter mappings are shared. Local normalization and private graph predictors preserve client-specific spatial and statistical variation.

This design addresses three constraints jointly: clients have different node sets and graph structures, useful traffic knowledge should be transferred across clients, and the shared information must remain compatible with each private predictor. Generic partial model sharing specifies which layers are shared, but does not provide this graph-dependent, node- and window-conditioned interface or the corresponding reconstruction–prediction training objective.

The added controls and analyses support this interpretation. The complete w/o FL model performs similarly to local AGCRN, while federation improves the same architecture, separating the benefit of cross-client learning from the mere addition of local modules. Prompt interventions and probes show that the learned prompt retains input-window information and that its node/window correspondence matters for forecasting; the BN ablation shows the role of local statistical calibration. The paper also provides a conditional analysis of the detached reconstruction–prediction weighting. These results evaluate the proposed sharing mechanism and its behavior, rather than only reporting the effect of combining standard components.

Reviewer_2qLpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes AutoFed, a personalized federated learning framework for traffic prediction that addresses data heterogeneity and varying graph sizes across clients. The framework combines a locally trained Personalized Predictor with a Federated Representor that generates adaptive prompts through an auto-encoder and a client-aligned adapter. Experiments on travel demand and traffic flow datasets demonstrate competitive or superior prediction performance compared with existing FL baselines.

Critical strengths

S1. The separation of shared feature-level modules and private graph-dependent components provides a reasonable approach to handling heterogeneous clients.

S2. The paper provides a theoretical complexity analysis to support the efficiency of the proposed framework.

S3. The experiments cover both travel demand prediction and traffic flow prediction using multiple real-world datasets and federated scenarios. AutoFed achieves competitive or superior performance compared with several federated learning baselines.

Critical weaknesses

W1. The paper emphasizes eliminating manual hyperparameter tuning as a key motivation. However, this appears to be a relatively minor practical issue, as existing automated hyperparameter optimization techniques could potentially address it. The authors should better justify why this problem requires a new federated learning framework rather than standard automated search strategies.

W2. The related work discussion is placed in the appendix rather than the main paper.

W3. The overall design lacks a sufficiently novel technical contribution. The proposed framework is too simple, which mainly integrates existing techniques, including AGCRN, auto-encoders, FedBN, and prompt-based conditioning. The methodological novelty beyond combining these established components is not sufficiently clear.

W4. The authors explicitly remove certain training strategies, such as curriculum learning, from the FedTPS baseline. Moreover, using a single hyperparameter configuration across all methods may disadvantage baselines that require method-specific tuning, potentially affecting the fairness of the comparisons.

W5. The paper does not sufficiently distinguish AutoFed from existing personalized federated learning approaches. More comparisons with recent and competitive PFL methods are necessary to establish the advantages of AutoFed. And the paper targets federated traffic prediction (FTP) task, more FTP works should be compared.

W6. Although AutoFed reduces communication costs compared with full-model federation, Table 3 shows that it communicates 175K parameters per round, substantially more than the 0.1K reported for FedTPS. Its computational overhead is also relatively high, raising questions about its overall efficiency.

W7. The convergence analysis assumes a fixed auxiliary loss weight, whereas the actual training algorithm employs an adaptive weighting strategy. Therefore, the theoretical guarantees do not directly apply to the proposed optimization procedure.

W8. The purpose of Figure 3 is not clearly explained. It is difficult to understand what the visualization is intended to demonstrate.

Questions for authors
  1. Is graph heterogeneity really a challenge? Existing methods presumably process each node separately through the model, making the model parameters independent of the number of nodes. Otherwise, how could these methods perform model aggregation or even train successfully?
6 comments
AuthorsOpenReview ↗

Response Part 1

Thank you for your careful review and constructive comments. We have added experiments and clarified the method, analysis, and experimental protocol. Our point-by-point responses follow.

W1: Motivation and hyperparameter optimization.

Thank you for your insightful comment. We clarified the motivation in the Introduction and Sections 2.2–2.3 and 3.2: AutoFed transfers shared knowledge to private graph predictors through learned prompts, while removing task-specific filter and repository-size choices. Avoiding manual tuning is one motivation; the contribution is this collaboration interface for heterogeneous traffic clients.

HPO evaluates candidate configurations through repeated training and validation. Traffic-specific choices such as filters, denoising settings, and repository size can require renewed searches when data or client partitions change. AutoFed replaces these choices with the learned AE/adapter pathway. Learning rate and hidden width remain configurable and can be optimized with HPO.

HPO selects configurations; AutoFed defines how clients share information across different graphs and statistics. Sharing only the representor's mappings also reduces communication.

W2: Related work in the main paper.

Thank you for your insightful comment. We moved the full Related Work into Section 2, before Methodology. Its three subsections cover Traffic Prediction, Federated Learning, and Prompt Learning, including recent traffic predictors and the distinctions from prior PFL approaches.

W3: Contribution beyond combining existing components.

Thank you for your insightful comment. We clarified the application-driven design in Sections 3.2–3.4 and supported it with complete local controls and prompt analyses. AutoFed addresses a practical problem in federated traffic prediction: clients need shared traffic knowledge while retaining their own road-network dependencies and traffic statistics. Shared prompt generation and local normalization provide that connection without exchanging graph predictors or selecting a discrete pattern repository.

The components serve this common purpose. Shared AE mappings learn temporal-feature transformations; private graph encoders incorporate local spatial relationships; prompts supply this knowledge to each private predictor for its current node and input window. Local BN calibrates the features entering the shared adapter. Together, these choices support different local graph sizes, limit communication, and retain client-specific forecasting models.

The complete local control distinguishes this collaboration mechanism from adding a local network. Here we show S3 and PEMS08, one setting per task, with all four metrics (mean ± standard deviation); complete results are in Section 4.2, Tables 1–2.

SettingMethodMAERMSEMSEMAPE (%)
S3 UberLocal AGCRN3.45±0.055.43±0.1033.69±1.5639.45±0.13
S3 Uberw/o FL3.44±0.035.42±0.0833.42±0.8139.32±0.20
S3 UberAutoFed3.36±0.045.27±0.0731.56±0.8938.82±0.49
S3 LyftLocal AGCRN2.00±0.043.06±0.069.61±0.3949.16±0.19
S3 Lyftw/o FL1.94±0.003.01±0.019.38±0.0750.21±0.30
S3 LyftAutoFed1.85±0.022.82±0.048.40±0.1950.33±0.29
PEMS08Local AGCRN18.65±0.3627.64±0.53790.42±23.9012.44±0.45
PEMS08w/o FL18.97±0.2127.73±0.20793.21±17.5413.33±0.33
PEMS08AutoFed16.83±0.1725.62±0.21684.63±14.1611.10±0.26

Without sharing, the added pathway performs similarly to local AGCRN; federation supplies the main gain. The BN ablation increases MAE across the reported settings. Appendix D provides complementary insights: adaptation reduces client-associated separation on TFP while retaining within-client structure, probes recover input-window statistics, and shuffling prompts damages forecasting despite preserving their values. Thus, the prompt retains useful traffic information whose correspondence to a particular node and window matters for prediction. W8 below gives the representative analysis tables; Section 4.2 and Appendix A.2 mark the sharing analysis and BN-ablation definition in blue.

AuthorsOpenReview ↗

Response Part 2

W4: Training strategies and baseline configurations.

Thank you for your insightful comment. We used each baseline's best-performing configuration reported in its paper, following the official implementation for method-specific settings. No method, including AutoFed, received additional hyperparameter tuning in our experiments. We removed the optional curriculum-learning schedule from the FedTPS implementation to compare the core methods without additional training enhancements.

Curriculum learning is an optional, general-purpose training heuristic included in the official FedTPS repository, rather than a component of the traffic-pattern-sharing method proposed in the paper. Such heuristics can also be applied to AutoFed and other compatible models to improve training. Combining multiple enhancements makes it difficult to distinguish gains from the proposed method from gains due to its training recipe. We therefore use a clean comparison without adding curriculum learning or stacking extra performance-enhancement heuristics, while retaining each method's core algorithm and method-specific hyperparameters.

Section 4.1 and Appendix A.2 specify the configuration sources and the curriculum-learning choice. Every method reports the test result from its best-validation round, and results are mean ± standard deviation over three runs. The common checkpoint-selection rule is marked in blue in Appendix A.2.

W5: PFL distinctions and recent/traffic-specific comparisons.

Thank you for your insightful comment. We added FedRMamba (WWW 2026) to Section 4.2, Tables 1–2. AutoFed has lower mean MAE in every TDP scenario-operator pair and on all three PEMS datasets. Sections 2.1–2.2 and 4.2 describe the architectural distinction: AutoFed explicitly models local spatial dependencies through adaptive graphs and transfers knowledge through prompts, while FedRMamba models interactions through a shared state-space scan over nodes.

Here we show S3 and PEMS08, one setting per task. Complete results are in Tables 1–2; entries use the same mean ± standard deviation convention.

SettingMethodMAERMSEMSEMAPE (%)
S3 UberFedRMamba3.63±0.005.68±0.0038.02±0.0744.41±0.03
S3 UberAutoFed3.36±0.045.27±0.0731.56±0.8938.82±0.49
S3 LyftFedRMamba2.01±0.002.93±0.019.66±0.0357.52±0.47
S3 LyftAutoFed1.85±0.022.82±0.048.40±0.1950.33±0.29
PEMS08FedRMamba19.28±0.4728.50±0.39907.41±39.5012.07±0.23
PEMS08AutoFed16.83±0.1725.62±0.21684.63±14.1611.10±0.26

Traffic nodes are coupled through spatial relationships. A scan over an ordered node sequence does not explicitly encode these graph neighborhoods. AutoFed learns an adaptive adjacency and uses graph convolution to aggregate relevant local nodes in both prompt generation and forecasting. This supplies a spatial inductive bias suited to traffic networks, while local BN accommodates different client distributions. This difference in spatial modeling offers a plausible explanation for AutoFed's stronger traffic-forecasting performance.

On PEMS both methods use the same single historical signal channel. On TDP, AutoFed uses nine historical channels and FedRMamba uses target history; Appendix A.2 specifies these inputs. Table 2 also retains FedRMamba's lower MSE on PEMS03. FedRMamba adds a recent PFL comparison; FedTPS and FedGCN provide the traffic-specific numerical comparisons, alongside the expanded discussion in Sections 2.2–2.3.

W6: Communication, computation, and overall efficiency.

Thank you for your insightful comment. We clarified the cost comparison in Section 4.3 and Appendix C. FedTPS communicates less per round because it shares a compact traffic-pattern repository. AutoFed shares learned prompt-generation mappings and achieves lower MAE across the regional TDP and PEMS settings. The relevant trade-off is the representation exchanged and the resulting accuracy, together with local computation.

Per-round costs on S1 with five Uber clients are:

MethodTime (s/round)Shared scalars/client/direction
FedAvg76.25.5M
FedTPS106.40.1K
AutoFed104.6175.0K

FedTPS's small repository explains its low communication cost, but also limits the capacity of the information shared across clients. The number of patterns is an explicit hyperparameter, creating a trade-off between repository capacity and payload. AutoFed replaces this discrete repository with an input-dependent neural representation, avoiding repository-size selection while improving prediction accuracy.

AuthorsOpenReview ↗

Response Part 3

Both payloads are small in absolute terms. AutoFed's 175K FP32 scalars occupy about 0.7 MB per direction per client per round, making the raw tensor transfer modest on links with adequate bandwidth. It still reduces the payload about 31-fold relative to full-model federation. Its extra graph encoder and reconstruction pathway add local work: time per round is higher than FedAvg's and close to FedTPS's. Section 4.3 marks the key clarification in blue, and Appendix C gives the counting and timing scope.

W7: Adaptive loss weighting and the theory.

Thank you for your insightful comment. We replaced the adaptive-rule argument with a fixed-batch potential analysis in Appendix E.1 and separated the fixed-weight SGD result in Appendix E.3. The new result establishes conditional local descent for the detached weighting, rather than convergence of the experimental adaptive Adam schedule. Section 3.5 states this distinction; key corrections are blue in Appendix E.

For a fixed client and batch, let a > 0 and b >= 0 be the differentiable prediction and reconstruction losses. The rule gives g=∇a+(b/a)∇bg=\nabla a+(b/a)\nabla b. For V=(a2+b2)/2V=(a^2+b^2)/2, we obtain ∇V=ag\nabla V=a g, so -g descends V away from stationarity. Appendix E.1 derives finite-step descent under smoothness, positivity, and step-size conditions.

Appendix E.2 gives a two-batch counterexample showing why the local result does not transfer directly through batch or client averaging. Appendix E.3 studies the fixed-weight personalized objective under full participation and one SGD step per client, retaining the stochastic noise floor.

W8: Purpose of the prompt visualization.

Thank you for your insightful comment. Our analyses show three useful properties of the learned prompts: the adapter reduces client-associated separation on TFP while largely preserving within-client geometry, the prompt retains information about the observed traffic window, and its correspondence to the correct node and window is important for forecasting. We made these findings explicit in Appendix D and clarified the original clustering plot, now Figure 4.

The clustering plot compares client profiles before and after adaptation: colors denote clients, and the horizontal axis denotes clusters. The profiles share a dominant cluster after adaptation. The original-space silhouette analysis quantifies the reduced client separation; for PEMS08 with four clients:

SettingBefore adapterAfter adapter
PEMS08 (4 clients)0.0980.013

The CKA analysis shows that within-client relationships are largely preserved on TFP, even as grouping by client weakens. TDP begins with weak client separation and undergoes a larger reorganization. The adapter thus reduces client-associated variation while retaining within-client structure on TFP (Appendix D.1).

Prompt replacement tests how the learned condition affects the fixed decoder. The table reports the relative increase in test MAE (%), averaged over clients, for TDP S1 and PEMS08. Complete results are in Appendix D.2, Table 7.

SettingZeroShuffleSwap-farSwap-near
TDP S1 (5 clients)43625142
PEMS08 (4 clients)137222119121

Zero removes the prompt; shuffle permutes its node/window assignment. Swap-far/near replace it with another client's mean prompt, choosing the donor by distance between client means. All replacements increase MAE. Shuffling preserves the collection of prompt values but is especially damaging on PEMS08, showing that a prompt's association with its input matters.

Linear probes examine the information retained in the prompt. The targets are the input window's mean and standard deviation; the control uses shuffled training labels. Entries are test R² (Appendix D.3, Table 8).

SettingMean probeMean controlStd. probeStd. control
TDP S1 (5 clients)0.140.000.200.00
PEMS08 (4 clients)0.790.000.350.00

The probe scores exceed the controls, with stronger recovery on TFP. Combined with the replacement results, this shows that the learned condition carries input-dependent traffic information that the decoder uses. Probes use a random split of representation rows; forecasting uses chronologically held-out windows. Appendix D marks the silhouette definition, node/window correspondence, and split distinction in blue.

Reviewer_54U9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AutoFed for personalized federated traffic prediction. Each client keeps its graph model private. Only a small prompt-generation module is shared between clients. This allows clients with different numbers of nodes to train together. The authors test the method on travel-demand and traffic-flow datasets. AutoFed generally performs well and uses less communication than full-model federation.

Critical strengths

The idea is simple and useful. Traffic clients may have different graph sizes, so their node-related parameters cannot be directly averaged. AutoFed avoids this problem by keeping these parameters local and sharing only a small prompt module.

The experiments cover several and enough datasets and baselines. The method usually achieves good prediction results with much lower communication than full-model federated learning.

Critical weaknesses

My main concern is that the paper does not show whether the improvement really comes from federated sharing.

The Local Training baseline does not use the full AutoFed architecture. A stronger comparison would train the complete AutoFed model locally without sharing the prompt generator. Without this experiment, it is hard to separate the benefit of federation from the benefit of the new architecture.

I am also not convinced by the scalability claim. For the PEMS datasets, the results for 2, 4, 6, 8, and 10 clients are averaged into one number. This hides how the method changes when the number of clients increases. The paper should report each setting separately.

Questions for authors
  1. What happens if each client trains the complete AutoFed model locally, without any federated aggregation?

  2. Can the authors report separate results for 2, 4, 6, 8, and 10 clients?

  3. Is the released code the exact version used for the paper? The code seems to put the prompt into the decoder’s initial hidden state, while the paper describes it as the decoder’s first token.

3 comments

Thank you for your careful review and constructive comments. We have added experiments and clarified the method, analysis, and experimental protocol. Our point-by-point responses follow.

W1 / Q1: Does the gain come from federated sharing?

Thank you for your insightful comment. We added the complete AutoFed model trained locally without sharing (w/o FL) to Section 4.2, Tables 1–2. Its performance is similar to the original local AGCRN baseline, whereas federation lowers MAE in every reported setting. This supports cross-client learning as the main source of the improvement.

Here we show one setting from each task, S3 and PEMS08; the complete results are in Table 1 for TDP and Table 2 for TFP. Local AGCRN is the original baseline, and w/o FL retains the complete AutoFed architecture and objective. Entries follow the mean ± standard deviation convention in those tables.

SettingMethodMAERMSEMSEMAPE (%)
S3 UberLocal AGCRN3.45±0.055.43±0.1033.69±1.5639.45±0.13
S3 Uberw/o FL3.44±0.035.42±0.0833.42±0.8139.32±0.20
S3 UberAutoFed3.36±0.045.27±0.0731.56±0.8938.82±0.49
S3 LyftLocal AGCRN2.00±0.043.06±0.069.61±0.3949.16±0.19
S3 Lyftw/o FL1.94±0.003.01±0.019.38±0.0750.21±0.30
S3 LyftAutoFed1.85±0.022.82±0.048.40±0.1950.33±0.29
PEMS08Local AGCRN18.65±0.3627.64±0.53790.42±23.9012.44±0.45
PEMS08w/o FL18.97±0.2127.73±0.20793.21±17.5413.33±0.33
PEMS08AutoFed16.83±0.1725.62±0.21684.63±14.1611.10±0.26

The local prompt pathway transforms the same observed history and provides little additional forecasting benefit on its own. The gain emerges when its mappings learn from other clients through federation. Appendix A.3 defines the control; the sharing analysis in Section 4.2 is marked in blue.

W2 / Q2: Separate results at 2, 4, 6, 8, and 10 clients.

Thank you for your insightful comment. We added the PEMS08 comparison at 2, 4, 6, 8, and 10 clients in Appendix A.4, Figure 2, with a reference in Section 4.2. Errors generally increase as the client count grows, particularly for the baselines, while AutoFed changes less and retains lower MAE, RMSE, and MSE at each count.

With the sensor set fixed, more clients mean smaller local graph partitions. This changes which spatial dependencies each client can model locally and can fragment useful cross-region relationships, helping explain the overall deterioration. Figure 2 provides all methods and metrics.

Q3: Released code and hidden-state conditioning.

Thank you for your insightful comment. We corrected the description in Sections 3.2 and 3.4 and Appendix B to match the released implementation: the prompt initializes the decoder hidden state. The model used in the experiments is unchanged.

For a recurrent decoder, a prefix token influences later predictions through the hidden state it produces. Hidden-state initialization introduces prompt information directly into that state, serving the same conditioning role. Our implementation concatenates the prompt with the encoder state. The correction specifies where conditioning occurs and preserves the method's core mechanism; the initialization description is marked in blue.

Reviewer_54U9OpenReview ↗

Response to Authors

The addon baseline experms has addressded my major concerns, supporting the benefits of fed learning. I have updated my rating from 2 to 3. However, the discrepancy between the Fig2 and Table3 is still with some doubts.

Thank you for raising your score and for the follow-up comment. Figure 2 and the main performance table come from separate training and evaluation runs. Small numerical differences can therefore arise from different random seeds and run-to-run variation. These differences do not change the qualitative conclusions or the observed performance trend. We hope this clarification address your concerns sufficiently.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers, AC, and PC for their time and effort in evaluating our submission. With one day remaining in the rebuttal period, we summarize AutoFed's contributions and the additions and clarifications made in response to the reviews.

Contributions

  • Input-dependent knowledge transfer across heterogeneous traffic graphs. AutoFed connects shared representation learning with private traffic predictors through a prompt generated for each node and history window. Shared auto-encoder and adapter mappings work with private graph encoders, allowing clients with different node sets and graph sizes to collaborate through a common interface.
  • Spatial and statistical personalization within the prompt interface. Adaptive local graphs learn traffic dependencies, while local normalization calibrates client-specific features before shared transformations. Together, these choices transfer knowledge while preserving local spatial structure and traffic statistics. The learned pathway replaces explicit pattern-extraction filters and a discrete pattern repository, reducing task-specific design and selection costs.
  • Prediction gains with reduced communication. Across the regional demand scenarios and all three PEMS flow datasets, AutoFed achieves lower error than the compared baselines. In the reported cost configuration, it reduces the shared parameter payload by a factor of approximately 31 relative to full-model federation. This provides a practical accuracy–communication trade-off for collaborative traffic prediction.

Main revisions during rebuttal

  • Isolated the benefit of federation. We added the complete AutoFed architecture trained without sharing, retaining the representor, adapter, and reconstruction objective. This control performs similarly to the original local AGCRN baseline, while federation lowers error across the evaluated settings. The results support cross-client learning as the main source of the improvement.
  • Expanded baseline and client-count comparisons. We added FedRMamba as a recent personalized federated forecasting baseline; AutoFed achieves lower error in every evaluated demand scenario–operator pair and on all three PEMS datasets. We also provided PEMS08 results separately by client count, showing stable AutoFed performance across the evaluated partitions.
  • Explained what the learned prompts capture and how they are used. Representation analysis shows reduced client-associated separation on flow prediction while largely preserving within-client structure. Linear probes recover observed-window statistics, and prompt interventions show that matching prompts to their nodes and input windows matters for forecasting. These analyses connect the prompt design to its predictive role.
  • Revised the analysis and presentation. We corrected the optimization analysis by deriving conditional local descent for the adaptive weighting on a fixed client and batch and separating it from the fixed-weight stochastic analysis. We clarified baseline configurations, checkpoint selection, repeated-run statistics, data partitions, AE dimensions, and the curriculum-learning choice. We also moved the full Related Work into the main paper, explained the application-driven design, aligned the conditioning description with the implementation, and clarified the communication–computation trade-off.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.