{"id":"ad108941-64bd-4463-81f2-13ca275432d1","arxiv_id":"2501.05934","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"An under-specified N-tier federated learning design with spatial encoding reports 75.62% and 89.52% accuracy on two geospatial datasets, but the encoding mechanism is never defined.","lead":"This paper proposes a multi-tier federated learning system that adds spatial encoding to predict greenhouse gas emissions and electric vehicle charging demand, reporting competitive accuracy in decentralized settings. A smart generalist might read it to see whether location-aware federated models can approximate centralized predictions without sharing raw geographic data.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Spatial encoding and spatial weight are never defined, and Algorithm 1's weighted aggregation contradicts Eq. (2), so the reported tier-2 accuracies cannot be attributed to the proposed mechanism.","rationale":"After reading the full text, the weakest point is indeed the spatial encoding and spatial weight. The paper's novelty claim rests on these operations, and Algorithm 1 is the only place they appear, but no equation or procedure defines them. This is not merely a missing detail; the pseudocode's line 13 conflicts with the formal aggregation in Eq. (2). The reported results could come from arbitrary weighting choices, and the centralized NN baseline outperforms the method in all tables, so there is no independent evidence that the encoded spatial mechanism works. I considered alternative concerns such as data leakage and the lack of error bars, but those are secondary; even if the experimental protocol were flawless, the method would still be under-specified. A simple ablation comparing the line-13 aggregation to plain FedAvg would clarify whether the undefined weights matter. The reader's verdict of REJECT is consistent with this reading, so no change is needed.","tokens_in":8471,"tokens_out":4313,"duration_ms":39648,"concrete_test":"Run an ablation on the same two datasets and tier-2 test splits: replace Algorithm 1 line 13's weighted bias sum with plain FedAvg (Eq. 2), keeping all other settings fixed. If plain FedAvg reproduces 75.62% and 89.52% within a small margin, then the undefined spatial weights are not load-bearing and the claimed advantage of spatial encoding is unsupported. If plain FedAvg performs substantially worse, the authors must release the exact spatial-weight formula and code; without it, the method remains irreproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that encoding spatial attributes in an N-tier FL system yields competitive aggregated models (75.62% gas, 89.52% energy at tier-2). The load-bearing novelty is the 'encode spatial data' and 'calculate spatial weight' steps, yet Section III provides no definition for either. Eq. (1) writes the local loss as L(w_i; D_i, s_i) but never states how s_i is computed or incorporated; Eq. (2) is plain FedAvg with equal weights. Algorithm 1 line 8 calls for 'calculate spatial weight,' and line 13 aggregates as `model += spatial weight * model bias`, which is neither FedAvg nor any weighted average with a defined normalization. Because the spatial weight is the only ingredient that distinguishes this method from standard FedAvg, and it is unspecified, the reported accuracies cannot be reproduced or attributed to spatial encoding. The paper's own tables show the centralized Neural Network outperforms the method, so the entire contribution rests on this undefined mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an N-tier federated learning (FL) framework in which client models first \"encode spatial data\" and then aggregate through multiple tiers, with tier-2 providing global or national predictions. The method is evaluated on two geospatial datasets: Canadian greenhouse gas emissions and New Brunswick electric-vehicle charging energy consumption. The authors report accuracies of 75.62% and 89.52% at tier-2, allegedly without training on data from that tier, and compare against centralized neural networks, ensemble learning, FedAvg, and weighted FedAvg. The central novelty is the spatial encoding and spatial-weight aggregation, but neither is formally defined, and the evaluation protocol is not specified in enough detail to verify the headline claims.","tokens_in":8699,"tokens_out":4505,"duration_ms":46044,"significance":"If the mechanism were precisely defined and the results were reproducible, the paper would offer a useful empirical data point on hierarchical FL for geospatial data: it shows that decentralized aggregation can approach centralized neural-network performance, and it explicitly compares several FL baselines. The authors also provide per-client prediction samples, which are descriptive and potentially valuable. However, the load-bearing components—\"spatial encoding\" and \"spatial weight\"—are undefined, and the empirical evaluation lacks error bars, statistical tests, and a description of how the \"unseen tier\" was held out. The manuscript ships no code, data splits, or hyperparameter details, so the claimed accuracies cannot currently be verified or attributed to the proposed mechanism. The significance is therefore not assessable at the present level of specification.","major_comments":[{"comment":"The operations that define the method are never specified. Eq. (1) uses a spatial encoding s_i and Algorithm 1 line 8 calls for \"encode spatial data\" and \"calculate spatial weight,\" but no equation, algorithm step, or textual description defines these operations. Since the spatial weight is the only ingredient distinguishing the method from standard FedAvg, the paper's reported accuracies cannot be attributed to the proposed mechanism. Additionally, Algorithm 1 line 13 aggregates as \"model += spatial weight * model bias,\" which is not the FedAvg aggregation stated in Eq. (2) and lacks any normalization or correctness argument. The method must be defined precisely before the experimental claims can be evaluated.","section":"Section III, Eq. (1), Algorithm 1"},{"comment":"The evaluation protocol for the central claim \"without having to train the model using the data constituted with the designated tier\" is not described. The paper does not state which data are held out from which tier, how the \"unseen tier\" split is created, or whether the spatial weights are computed using validation data from that tier. This leaves open the possibility of information leakage between spatial-weight estimation and evaluation. The accuracy tables report single runs with no error bars or statistical significance tests; the headline tier-2 gap between the proposed method (75.62%) and the neural network (76.20%) is within the range of random variation for a single run. The authors should provide the exact data split, multiple seeds, and standard deviations.","section":"Section IV, Section V, Tables I and III"},{"comment":"The text states, \"Input features included geo-location, reference dates, target values, and other relevant factors.\" If target values are indeed included as input features, the accuracy results are invalidated by label leakage. The authors must clarify whether the target variable was used as a feature, and, if so, remove it and rerun the experiments; if the wording is an error, it must be corrected because it directly affects the credibility of every reported accuracy.","section":"Section IV, input features"},{"comment":"The headline results are not traceable in the tables. Tables I and III are labeled \"Tier-1\" and list server-level accuracies, while Figures 2 and 3 are labeled \"tier-2\"; the claimed tier-2 accuracies of 75.62% and 89.52% do not appear in any table. The reader cannot verify whether the tier-1 and tier-2 results are computed on the same model family, the same data splits, or the same aggregation procedure. Please provide full numeric tables for every tier and every experiment, including per-tier baseline numbers.","section":"Tables I and III; Figures 2 and 3"}],"minor_comments":[{"comment":"The control flow of Algorithm 1 is unclear: \"while Num Tiers\" does not specify a loop over tiers, and \"if Num Tiers=0\" is used as the client-level condition. This should be rewritten as a proper loop over tiers and clients.","section":"Algorithm 1"},{"comment":"Section IV says \"simple two-layer neural networks were deployed,\" but Figure 1 depicts three fully connected layers (FC1, FC2, FC3). Please reconcile the architecture description and the figure.","section":"Section IV, Figure 1"},{"comment":"The sentence \"FedAvg Weighted method had the lowest (9.47%)\" in the gas-emission discussion does not correspond to any value in Tables I or III; please add the exact table or figure and state which tier and server this value refers to.","section":"Section V"},{"comment":"Tables II and IV show only five predictions per client and no per-client accuracy or aggregate error measure; please include the per-client accuracy and the class distribution, since the reported global accuracy cannot be checked from these samples.","section":"Tables II and IV"},{"comment":"The NB power dataset is cited as both [27] and [29]; please unify the references and clarify which source corresponds to the dataset used in the experiments.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is not close to publication in its current form. The proposed method is underspecified at its core, the evaluation protocol does not support the headline 'unseen tier' claim, and the reported accuracies are not reproducible from the description. These are load-bearing issues that require defining the method, rerunning the experiments with proper splits and error bars, and, in all likelihood, correcting the feature description. Given the scope of the missing material, I do not see how a normal revision cycle can resolve them without substantially rewriting the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes an N-tier federated learning scheme with 'spatial encoding' at the client tier, but that encoding is never defined. The only aggregation equation, Eq. (2), is plain FedAvg with equal weights, and Algorithm 1 line 13 aggregates as `model += spatial weight * model bias`, which is neither FedAvg nor any weighted average with stated normalization. So the central mechanism is unverifiable, and the claimed accuracies (75.62% on gas, 89.52% on energy) cannot be attributed to spatial encoding.\n\nThat said, the paper does useful applied work: it runs hierarchical FL on two real geospatial datasets, compares against neural net, ensemble, FedAvg, and weighted FedAvg baselines, and reports per-client predicted vs. actual classes. The idea of assessing a tier-2 model without training on that tier's data is a reasonable question, even if the data split is not described.\n\nThe soft spots are load-bearing. The 'encode spatial data' and 'calculate spatial weight' steps appear in Algorithm 1 but have no definition anywhere in the paper; Eq. (1) includes s_i but never says how it is computed or used. The evaluation reports single runs without error bars or statistical tests, and no code or data is provided. The paper itself states the centralized neural network outperforms the proposed method on every server, so the contribution reduces to an undefined weighting step. The stress-test concern about internal contradiction is correct: Eq. (2) and Algorithm 1 cannot both be the aggregation rule.\n\nWho is this for? Readers interested in a preliminary application of hierarchical FL to spatial emission/energy data, and as a cautionary example of a paper whose claimed novelty is not pinned down. It is not ready for publication, but the topic and datasets are real, and a revision that defines the encoding/weight, corrects the algorithm, adds error bars, and clarifies the data split could make it worth another look.\n\nMy recommendation: send it to peer review rather than desk reject. A serious referee will quickly identify the missing definitions and the algorithm inconsistency, and the authors need that feedback. But with the method undefined, my own verdict is skeptical. In short, the paper deserves a serious referee, but it is not close to acceptance as written.","headline":"The paper's 'spatial encoding' is never defined, and its aggregation algorithm contradicts its own FedAvg equation, so the reported tier-2 accuracies cannot be attributed to the proposed mechanism.","tokens_in":9185,"tokens_out":3727,"would_cite":false,"duration_ms":34183,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Encoded multi-tier federated learning predicts unseen spatial tiers at 75.62% and 89.52% accuracy.","keywords":["federated learning","spatial data","multi-tier aggregation","spatial encoding","geospatial prediction","FedAvg","edge computing","model evaluation"],"falsifier":"A concrete check: rerun both experiments with an explicit published definition of spatial encoding and spatial weight, for instance coordinates as augmented features and inverse-distance weighting. If the tier-2 accuracies fall back to the FedAvg baselines, or if the spatial weights only reach the reported levels when computed using the held-out tier's data, then the central claim is not supported.","tokens_in":8308,"feed_emoji":"🌍","tokens_out":7221,"duration_ms":67779,"temperature":0.7,"pith_summary":"This paper proposes an N-tier federated learning architecture for geospatial prediction in which each client encodes the spatial information in its local data, trains a small neural network, and sends its model bias and a spatial weight to higher tiers for Federated Averaging. The central claim is that this encoded multi-tier aggregation produces models that predict well on spatial granularities they never trained on: 75.62% accuracy for greenhouse-gas emission classes and 89.52% accuracy for electric-vehicle energy consumption classes at tier-2. The authors position the contribution as extending federated learning to spatial data, where clients are provinces, territories, cities, or charging stations and each tier corresponds to a different spatial scale. A sympathetic reader would care because, if the claim holds, decentralized spatial prediction can approach the accuracy of a centrally trained neural network while keeping data local.","feed_headline":"N-tier federated learning predicts unseen spatial data at 89.5%","feed_subtitle":"Spatial encoding plus tiered FedAvg reaches 89.5% on energy, 75.6% on emissions.","key_machinery":"The carrying mechanism is Algorithm 1's spatial-encoded N-tier pipeline. At the client tier, each client encodes its spatial data, trains a two-layer neural network with the Adam optimizer, and computes a spatial weight; the local model bias and spatial weight are then sent upward. At each higher tier, a server aggregates the received models by combining each model bias with its spatial weight before applying Federated Averaging (FedAvg), so the spatial weight is what distinguishes this procedure from plain FedAvg. Tiers are organized by spatial granularity, with tier-1 representing regional or city-level aggregates and tier-2 the global level.","core_discovery":"On the paper's own terms, the discovery is that adding a spatial-encoding step at the client tier and then aggregating through multiple FedAvg tiers yields global models that are competitive with a centralized neural network even when the evaluated tier's data was not used in training at that tier. For the gas emission dataset the tier-2 encoded model reaches 75.62% accuracy against 76.20% for the neural network, and for the energy dataset it reaches 89.52% against 98.55%, while outperforming vanilla FedAvg, weighted FedAvg, and ensemble learning on most tier-1 servers. The paper also reports per-client prediction tables showing that the localized models match the actual class labels closely in most clients, which it treats as evidence that multi-tier aggregation can provide region-specific models rather than only one global model.","pith_inferences":["Beyond the paper: if the spatial encoding is simply a feature augmentation such as appending coordinates to the inputs, then the reported gains are a feature-engineering effect and should be reproducible by any federated learner using the same features, without needing a new aggregation rule.","Beyond the paper: a natural specification to test is inverse-distance weighting among clients within a tier for the spatial weight; the method's edge over plain FedAvg should persist under that explicit choice if the spatial mechanism is real.","Beyond the paper: applying the same tiered encoding to spatiotemporal data could reveal whether multi-tier aggregation acts as a regularizer on non-IID spatial distributions, a question the current two-dataset setup cannot separate from simple feature effects."],"forward_implications":["Tier-2 encoded FL reaches 75.62% on gas emissions and 89.52% on energy consumption without training on the designated tier's data, so hierarchical aggregation can generalize to unseen spatial granularities.","On tier-1 servers the proposed method beats vanilla FedAvg, weighted FedAvg, and ensemble learning on most servers for both datasets, indicating that the spatial encoding helps under uneven data distributions.","The per-client prediction tables show localized models matching actual labels in most clients, supporting the use of multi-tier FL to obtain region-specific models alongside a global model.","The higher accuracy on the more evenly distributed energy dataset suggests that data distribution, not architecture alone, drives multi-tier spatial FL performance."],"supporting_citations":[{"why":"supplies the greenhouse gas emissions dataset that drives Experiment 1","marker":"[11]"},{"why":"identified as the NB Power energy consumption dataset for Experiment 2","marker":"[27]"},{"why":"cited as the source of the New Brunswick electric-vehicle charging events used in Experiment 2","marker":"[29]"},{"why":"provides the Federated Averaging rule that carries each tier's aggregation","marker":"[28]"},{"why":"the Adam optimizer used for local model training on every client","marker":"[30]"},{"why":"defines the ensemble learning baseline the experiments compare against","marker":"[31]"},{"why":"defines the weighted FedAvg baseline that the proposed method must outperform","marker":"[32]"}],"fun_headline_variants":["Multi-tier federated learning with spatial encoding reaches 89.5%","N-tier FL predicts unseen spatial data at 89.5%","Spatial encoding and tiered aggregation yield 89.5% on unseen data","Encoded spatial attributes improve multi-tier FL accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the operations 'encode spatial data' and 'calculate spatial weight' in Algorithm 1 are well-defined procedures that preserve spatial signal; no equation or pseudocode defines them, and the reported accuracies depend on those unspecified steps.","fun_headline_variants_meta":{"raw":{"variants":["Multi-tier federated learning with spatial encoding reaches 89.5%","N-tier FL predicts unseen spatial data at 89.5%","Spatial encoding and tiered aggregation yield 89.5% on unseen data","Encoded spatial attributes improve multi-tier FL accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000861,"raw_usage":{"total_tokens":3728,"prompt_tokens":927,"completion_tokens":2801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2725}},"tokens_in":543,"tokens_out":2801,"duration_ms":17602,"temperature":1.0,"reasoning_tokens":2725,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:05:37.334392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: rerun both experiments with an explicit published definition of spatial encoding and spatial weight, for instance coordinates as augmented features and inverse-distance weighting. If the tier-2 accuracies fall back to the FedAvg baselines, or if the spatial weights only reach the reported levels when computed using the held-out tier's data, then the central claim is not supported.","supporting_citations":[{"cited_title":"Physical flow account for greenhouse gas emissions","cited_arxiv_id":null,"evidence_quote":"supplies the greenhouse gas emissions dataset that drives Experiment 1"},{"cited_title":"Discovering ev recharging patterns through an automated analytical workflow,","cited_arxiv_id":null,"evidence_quote":"identified as the NB Power energy consumption dataset for Experiment 2"},{"cited_title":"Evaluating multi-global server architecture for federated learning,","cited_arxiv_id":null,"evidence_quote":"cited as the source of the New Brunswick electric-vehicle charging events used in Experiment 2"},{"cited_title":"Decentralized federated averaging,","cited_arxiv_id":null,"evidence_quote":"provides the Federated Averaging rule that carries each tier's aggregation"},{"cited_title":"A survey on ensemble learning,","cited_arxiv_id":null,"evidence_quote":"defines the ensemble learning baseline the experiments compare against"},{"cited_title":"Revisiting weighted aggregation in federated learning with neural networks,","cited_arxiv_id":null,"evidence_quote":"defines the weighted FedAvg baseline that the proposed method must outperform"}],"review_version":1}