{"id":"c08b7694-a4ad-43e7-921a-dfb335061d4c","arxiv_id":"2412.06815","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FBTTR applies FedAvg-style averaging to block-term tensor regression and reports competitive accuracy on ECoG finger decoding and heart disease prediction tasks.","lead":"This paper introduces FBTTR, a federated learning version of the block-term tensor regression method, allowing hospitals to train shared models without pooling patient data. It reports accuracy on brain signal finger decoding and heart disease prediction, claiming results comparable or superior to centralized models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FBTTR's FedAvg aggregation of BTTR factors has no alignment step; permuted or sign-flipped components across clients make the averaged 'global' model an incoherent tensor decomposition, undermining the central claim.","rationale":"I read the paper as claiming that FBTTR is a valid federated extension of BTTR, i.e., that averaging local BTTR parameters gives a global tensor-regression model with privacy-preserving collaborative training and predictive performance comparable to centralized training. For that claim to hold, the block-term decompositions from different clients must be mutually identifiable up to the same permutation, sign, and scaling convention. The paper provides no identifiability result and no alignment mechanism; Section 2.3 only matches dimensions. This is the weakest link in the argument. The reader's weakest-assumption identifies exactly this issue, so I agree. The concern is not merely theoretical: ACE is an automatic, data-dependent procedure, and component order is not a continuous quantity that FedAvg can average over. I would not need to rely on the Table 5 discrepancy (FBTTR 0.872/0.772 vs BTTR 0.874/0.783 on the heart-disease data) to reject, although that discrepancy independently contradicts the advertised 'outperforms centralized BTTR' claim. A concrete synthetic test with permutation diagnostics would settle whether the aggregation is valid. Since the paper currently contains no such diagnostic and no theoretical guarantee, the reader's REJECT verdict is appropriate; my read does not change it.","tokens_in":18432,"tokens_out":8074,"duration_ms":73958,"concrete_test":"Re-run FBTTR on a synthetic dataset generated from a known 3-block BTTR model split across two clients. Instrument Figure 2 to record, before averaging, the ordered columns of each client's factor matrices P^(n)_k and core tensors G_k. Compute the absolute cross-correlation matrix between client factor columns and solve the linear assignment problem to detect the optimal permutation and sign alignment. Compare the test error of the naive FedAvg global model with (a) each local model and (b) a global model formed after applying the optimal permutation and sign alignment. If the optimal assignment is not the identity or the aligned average outperforms the naive average by a non-negligible margin, the aggregation in Section 2.3 lacks the alignment needed to support the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the FedAvg aggregation in Section 2.3 (Figure 2) produces a valid global BTTR model. That requires the block components learned by different clients to be aligned, since block-term decompositions are unique only up to permutation, sign, and scaling of the factors. The paper only checks that the dimensions of G and P^(n) match across clients; it never checks or enforces that client k's block 1 corresponds to client k' 's block 1, nor does it align signs or scales before averaging. With ACE's automatic rank and sparsity selection run independently on each client, different orderings and sign conventions are expected, especially under non-IID healthcare data. If components are permuted, the averaged parameter tensor does not represent any single block-term regression, so the global predictions and the reported AUC and C-index gains cannot be attributed to a coherent decentralized tensor model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Federated Block-Term Tensor Regression (FBTTR), a federated extension of BTTR for decentralized healthcare analysis. Each client trains a local BTTR model; the server selects common component dimensions from per-client ACE results and averages the local factor matrices and core tensors using a FedAvg-style update. The method is evaluated on BCI Competition IV finger-movement decoding and on two Flamby clinical datasets (Fed-Heart-Disease and Fed-TCGA-BRCA), comparing against centralized BTTR, eBTTR, HOPLS, non-multilinear baselines, and standard federated learning algorithms. The central claims are that FBTTR preserves privacy while achieving predictive performance comparable to or better than centralized BTTR, and that it outperforms non-multilinear models in the BCI case study.","tokens_in":18673,"tokens_out":7853,"duration_ms":60685,"significance":"If the claims held, FBTTR would be a useful contribution at the intersection of tensor regression and federated learning, and the paper deserves credit for making the implementation open-source and for re-implementing comparison baselines on public datasets. However, the central comparative claims are undermined by a direct contradiction between the abstract and Table 5, by a statistical test that cannot reach significance with five paired blocks, and by the absence of any alignment step in the averaging of non-identifiable tensor decompositions. Because these issues affect the paper's main conclusions, the significance of the contribution cannot be assessed from the present manuscript.","major_comments":[{"comment":"The central claim that FBTTR outperforms centralized BTTR on Fed-Heart-Disease is directly contradicted by Table 5: FBTTR obtains ROC-AUC 0.872 ± 0.02 and accuracy 0.772 ± 0.02, while the BTTR row reports 0.874 ± 0.03 and 0.783 ± 0.04. The numbers quoted in the abstract (0.872 vs 0.812 and 0.772 vs 0.753) correspond to the FL4E 'Centralized' row, not to the BTTR row. The text in §3.2.2 even acknowledges a 'slight accuracy drop' in the federated setting, so the abstract and conclusion overstate the result.","section":"Abstract, §3.2.2, Table 5"},{"comment":"The manuscript states that the test data were split into 5 non-overlapping blocks and that a two-tailed Wilcoxon signed-rank test was used with a significance threshold of p<0.05. With n=5 paired blocks, the smallest attainable two-tailed p-value for a Wilcoxon signed-rank test is 2/32 = 0.0625. Therefore none of the 'statistically significant' claims in Sections 3.1.2 and 3.2.2, including the bolded comparisons in Tables 5 and 6, can be supported by the stated test at the 0.05 significance level.","section":"§3.1.1 and §3.2.1"},{"comment":"The FedAvg-style aggregation averages the local factor matrices P(n)_kb and core tensors G(X)_kb entry-wise across clients. Block-term decompositions are unique only up to permutation of the blocks and scaling/sign (or more general invertible) transformations within each block. The algorithm checks only that the dimensions of the factor matrices are equal across clients; it does not align block orderings or sign/scale conventions before averaging. If local components are permuted or sign-flipped, the averaged parameter tensor does not correspond to a valid block-term regression, so the reported predictive performance cannot be attributed to a coherent decentralized tensor model. The limitations paragraph in Section 4 mentions computational overhead and heterogeneous-data degradation but not this identifiability problem.","section":"§2.3, Figure 2"},{"comment":"The claim that FBTTR 'performs statistically significantly better than BTTR' on Fed-TCGA-BRCA is unsupported for two reasons. First, with 5 paired blocks the two-tailed Wilcoxon test cannot yield p<0.05, as noted above. Second, the table shows local Client 4 achieving C-index 0.851 ± 0.078, which exceeds FBTTR's 0.775 ± 0.01, and Hybrid FedAdagrad reaching 0.776 ± 0.036, so the absolute comparison is not uniformly in FBTTR's favor.","section":"§3.2.2, Table 6"}],"minor_comments":[{"comment":"The method name is written inconsistently: Tables 2 and 3 use 'F-BTTR', Table 4 uses 'FBTTR', and the text and other tables use 'FBTTR'; please standardize the notation.","section":"Throughout"},{"comment":"The abstract and Section 4 describe the heart-disease result as outperforming centralized BTTR, while Section 3.2.2 states there is a slight, non-significant accuracy drop and 'practically no difference' in ROC-AUC; these statements need to be reconciled.","section":"Abstract, §3.2.2, §4"},{"comment":"Line 5 of Figure 2 instructs the server to 'Find SNRk and τk for client k such that all clients have the same dimensions' without explaining how a single set of dimensions is chosen from the per-client ACE results; please specify the selection rule.","section":"§2.3, Figure 2"},{"comment":"The BIC formula is written as log(||C - ...||_F / s) + log(s)/s * DF; the placement of s is ambiguous, and the second term would normally be log(s) multiplied by DF rather than divided by s. Please clarify the definition and cite the exact BIC form used.","section":"Equation (2)"},{"comment":"Figure 7 is described as showing model parameters for 'the first 30 blocks', but the number of blocks K is selected by cross-validation; please explain how 30 relates to the chosen K and why 30 was chosen for display.","section":"§3.1.2, Figure 7"},{"comment":"There are numerous typographical errors, including 'seperate', 'F ederation', 'FedA VG', and 'scenarions'; a careful proofread is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper has a serious internal contradiction between the abstract and Table 5, an underpowered statistical procedure that invalidates the significance claims, and a fundamental identifiability gap in the aggregation step. These are not merely presentational issues; correcting them would require new methodological work (an alignment step) and a full re-run of the experiments with valid statistics. I recommend rejection, although the open-source implementation and the use of public benchmarks are positive features that could form the basis of a future revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes Federated Block-Term Tensor Regression (FBTTR), which is exactly what it sounds like: BTTR with FedAvg. That extension is genuinely new, and the authors put together a real evaluation on two healthcare datasets (BCI finger decoding and Flamby heart disease/Tcga-Brca), re-implementing several baselines. The dimension-matching step, where the server forces the same number of components across clients, is a sensible practical move. For the BCI case, the results show FBTTR roughly on par with eBTTR and a bit better than HOPLS, which is a reasonable outcome.\n\nBut the paper has three serious problems. First, the abstract claims FBTTR outperforms centralized BTTR on the heart disease dataset, but Table 5 shows the opposite: the row labeled BTTR gets ROC-AUC 0.874 and accuracy 0.783 versus FBTTR's 0.872 and 0.772. The abstract's numbers (0.872 vs 0.812 for the centralized model) also don't line up with the table's BTTR row. This inconsistency is hard to excuse. Second, the authors use a two-tailed Wilcoxon signed-rank test on five non-overlapping blocks. With five pairs, the smallest possible two-tailed p-value is 0.0625, so any claim of p<0.05 is mathematically impossible. Third, and most conceptually, the FedAvg aggregation in Figure 2 averages factor matrices and cores without aligning the block ordering or the sign/scale of components across clients. Block-term decompositions are only defined up to permutation and scalar ambiguities per block. The paper checks that the dimensions match, but not that block 1 on client A corresponds to block 1 on client B. If components are misaligned, the averaged tensor doesn't represent any coherent BTTR model. Showing stable parameters for one subject doesn't address this in a heterogeneous federated setting. The stress-test note on alignment holds up.\n\nThere are also smaller issues: the promised open-source code isn't linked, and the intro has boilerplate text on vertical FL that reads like it belongs elsewhere.\n\nThe paper is for researchers working on federated tensor methods. They'll find the idea worth building on, but not the current results. The paper deserves a serious referee because the core question is legitimate and the experimental setup is substantial, but only after major revision. If the authors fix the statistical test, reconcile Table 5 with the abstract, and add an alignment mechanism with a theoretical justification, the method could be a useful contribution.","headline":"FBTTR is a plausible federated extension of BTTR, but the empirical claims are undercut by internal contradictions and the aggregation ignores tensor component alignment.","tokens_in":19148,"tokens_out":5229,"would_cite":false,"duration_ms":45708,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Federated Block-Term Tensor Regression lets institutions train shared tensor models without sharing patient data, matching or beating centralized performance on healthcare benchmarks.","keywords":["federated learning","block-term tensor regression","tensor decomposition","privacy-preserving machine learning","brain-computer interface","heart disease prediction","FedAvg","healthcare analytics"],"falsifier":"Train FBTTR on any of the paper's datasets, then before the aggregation step randomly permute the block indices and flip the signs of one client's core tensor and factor matrices. If the resulting averaged global model loses a large share of its prediction accuracy, the method's reported success depends on an implicit component alignment that the algorithm does not enforce. If the accuracy is unchanged, the averaging is robust to the decomposition ambiguity and the concern is moot.","tokens_in":18260,"feed_emoji":"🏥","tokens_out":5909,"duration_ms":50240,"temperature":0.7,"pith_summary":"The paper introduces Federated Block-Term Tensor Regression (FBTTR), a version of the Block-Term Tensor Regression (BTTR) algorithm that can be trained across multiple institutions without sharing raw data. Using FedAvg-style averaging of the tensor decomposition parameters across clients, FBTTR aims to give hospitals and research sites a privacy-preserving way to build high-dimensional predictive models collaboratively. The authors claim that FBTTR matches or exceeds centralized BTTR on two healthcare tasks: decoding finger movements from ECoG signals and predicting heart disease. If correct, this makes multilinear tensor regression usable in real-world federated healthcare collaborations. The paper reports, for example, a heart-disease ROC-AUC of 0.872 for FBTTR versus 0.812 for the centralized model.","feed_headline":"Federated tensor regression keeps pace with centralized models","feed_subtitle":"FBTTR matches or beats centralized BTTR on heart-disease and brain-signal tasks.","key_machinery":"The carrying mechanism is the block-term tensor decomposition at the heart of BTTR: the input tensor is written as a sum of blocks, each a Tucker decomposition composed of a core tensor $\\mathcal{G}_k$ and factor matrices $\\mathbf{P}^{(n)}_k$, with latent scores $t_k$ connected to the response through a regression coefficient. FBTTR separates Automatic Component Extraction (ACE) from fitting: each client runs ACE locally, the server picks shared SNR and $\\tau$ thresholds so that all clients produce factor matrices with the same dimensions, and then FedAvg averages the corresponding parameters across clients, weighted by local sample sizes. The dimension-synchronization step is what makes the averaged tensor model well-defined at the matrix level.","core_discovery":"On the paper's own terms, the central discovery is that a block-term tensor regression model can be federated by averaging the local factor matrices, core tensors, and regression weights with FedAvg, after synchronizing the component dimensions across clients through a federated version of Automatic Component Extraction. The resulting global model is claimed to achieve predictive performance comparable to, and sometimes better than, a centralized BTTR trained on pooled data. Evidence includes heart-disease prediction (Fed-Heart-Disease: ROC-AUC 0.872 ± 0.02 and accuracy 0.772 ± 0.02 vs. 0.812 ± 0.003 and 0.753 ± 0.007 for centralized BTTR) and finger-movement decoding (for subject 3's thumb, Pearson correlation 0.76 ± .05 vs. 0.71 ± .05 for centralized BTTR). In the breast-cancer survival dataset, FBTTR also reports a higher C-index than centralized BTTR (0.775 vs. 0.737), which the authors attribute to the federated setting leveraging data variance across centers.","pith_inferences":["The paper does not address the non-uniqueness of tensor decompositions when averaging local models. A natural follow-up would be to test whether aligning components across clients (e.g., by matching factors via correlation or Procrustes rotation) changes performance; if it does, the reported gains may depend on the data happening to align rather than on the algorithm itself.","The 'privacy' claim here is about not sharing raw data; the paper does not implement differential privacy or secure aggregation, so a direct extension is to combine FBTTR with those mechanisms to assert formal privacy guarantees.","The reported superiority of FBTTR over centralized BTTR on the cancer dataset may reflect the regularization or ensembling effect of averaging across heterogeneous sites rather than a property of tensor regression; a test would be to compare against an ensemble of locally trained BTTR models without any aggregation.","FBTTR currently targets horizontal FL with shared features; extending to vertical FL, where sites hold different features of the same patients, would require a different aggregation schema and is unaddressed."],"forward_implications":["If FBTTR works as claimed, hospitals and research centers can jointly fit BTTR-style tensor regressions without moving patient data, which would make high-dimensional predictive modeling available to multi-site healthcare collaborations bound by privacy rules.","The federated-hybrid result suggests that mixed setups, where some sites contribute centrally and others federate, may also be viable, pointing toward flexible deployments that respect local data-sharing policies.","FBTTR's performance relative to standard federated learners like FedAvg, FedProx, and adaptive optimizers indicates that tensor structure in the data is being exploited by the federated model, not just averaged away.","The method's success on survival data (TCGA-BRCA) suggests federated tensor regression can generalize beyond binary classification to time-to-event outcomes, if the tensor representation is available."],"supporting_citations":[{"why":"Introduces BTTR with fast automatic component extraction, the local regression model that FBTTR federates.","marker":"[6]"},{"why":"Provides the sparse Tucker decomposition (PSTD) that underlies the F-mPSTD step inside ACE, controlling core sparsity.","marker":"[43]"},{"why":"Defines Federated Averaging (FedAvg), the aggregation rule FBTTR uses for averaging model parameters.","marker":"[10]"},{"why":"Supplies the Fed-Heart-Disease and Fed-TCGA-BRCA clinical datasets used for the healthcare case studies.","marker":"[52]"},{"why":"Provides the FL4E framework and the fully-federated, hybrid, and centralized experimental scenarios used for evaluation.","marker":"[53]"},{"why":"The amplitude-modulation linear regression baseline whose reimplementation FBTTR is compared against in finger decoding.","marker":"[46]"},{"why":"The deep-learning (CNN/LSTM) baselines for finger decoding that FBTTR claims to outperform.","marker":"[47]"}],"fun_headline_variants":["Federated tensor regression matches centralized health models","Privacy-preserving tensor regression rivals central in healthcare","FBTTR outperforms central models on heart and brain data","Decentralized tensor regression: privacy without accuracy loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a simple weighted average of the clients' factor matrices, core tensors, and regression weights is a valid global BTTR model, which in turn assumes the local tensor decompositions are aligned in component order and sign; the paper only checks that the matrix dimensions match, not that the components correspond.","fun_headline_variants_meta":{"raw":{"variants":["Federated tensor regression matches centralized health models","Privacy-preserving tensor regression rivals central in healthcare","FBTTR outperforms central models on heart and brain data","Decentralized tensor regression: privacy without accuracy loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1535,"prompt_tokens":1083,"completion_tokens":452,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":389}},"tokens_in":699,"tokens_out":452,"duration_ms":4751,"temperature":1.0,"reasoning_tokens":389,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:21:47.842980+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FBTTR on any of the paper's datasets, then before the aggregation step randomly permute the block indices and flip the signs of one client's core tensor and factor matrices. If the resulting averaged global model loses a large share of its prediction accuracy, the method's reported success depends on an implicit component alignment that the algorithm does not enforce. If the accuracy is unchanged, the averaging is robust to the decomposition ambiguity and the concern is moot.","supporting_citations":[{"cited_title":"For each SNR value, the associated optimal τ is computed as τ ∗ = argmin τ BIC (τ, SNR)","cited_arxiv_id":null,"evidence_quote":"Introduces BTTR with fast automatic component extraction, the local regression model that FBTTR federates."},{"cited_title":"Signal quality of simultaneously recorded invasive and non-invasive eeg","cited_arxiv_id":null,"evidence_quote":"Provides the sparse Tucker decomposition (PSTD) that underlies the F-mPSTD step inside ACE, controlling core sparsity."},{"cited_title":"Walking naturally after spinal cord injury using a brain–spine interface","cited_arxiv_id":null,"evidence_quote":"Defines Federated Averaging (FedAvg), the aggregation rule FBTTR uses for averaging model parameters."},{"cited_title":"Decoding of finger trajectory from ecog using deep learning","cited_arxiv_id":null,"evidence_quote":"Supplies the Fed-Heart-Disease and Fed-TCGA-BRCA clinical datasets used for the healthcare case studies."},{"cited_title":"Prediction of finger flexion: 4th brain-computer interface data competition","cited_arxiv_id":null,"evidence_quote":"Provides the FL4E framework and the fully-federated, hybrid, and centralized experimental scenarios used for evaluation."},{"cited_title":"Give me a sign: decoding four complex hand gestures based on high-density ecog","cited_arxiv_id":null,"evidence_quote":"The amplitude-modulation linear regression baseline whose reimplementation FBTTR is compared against in finger decoding."},{"cited_title":"Gesture decoding using ecog signals from human sensorimotor cortex: a pilot study","cited_arxiv_id":null,"evidence_quote":"The deep-learning (CNN/LSTM) baselines for finger decoding that FBTTR claims to outperform."}],"review_version":1}