Pith. sign in

REVIEW 3 major objections 5 minor 18 references

CoxNTF: A New Approach for Joint Clustering and Prediction in Survival Analysis

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CoxNTF compresses baseline covariates into interpretable survival-linked factors without losing Coxnet's predictive accuracy.

desk verdict A genuinely new survival-weighted tensor factorization with an underspecified evaluation protocol that may leak test-set survival information into the meta-scores. read the letter →

arxiv 2506.06411 v2 pith:MTPQTUXP submitted 2025-06-06 cs.LG

Paul Fogel (1) , Christophe Geissler (1) , George Luta (2) ((1) Data Services , Forvis Mazars , Levallois , France , (2) Department of Biostatistics , Bioinformatics
show 5 more authors
Biomathematics Georgetown University Medical Center Washington DC USA)
This is my paper
classification cs.LG
keywords survivalanalysisnon-negativetensorfactorizationCoxnetlatentrepresentationclusteringcensoringconcordanceindexfeatureredundancy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Survival models usually need original covariates, and unsupervised latent representations like NMF lose survival information. CoxNTF instead builds a three-dimensional tensor whose third axis is time periods, weights each covariate entry by the Coxnet-estimated probability that the event occurs in that period, and factorizes the tensor non-negatively into patient meta-scores, covariate patterns, and temporal patterns. The paper claims that feeding these meta-scores to Coxnet gives survival predictions as accurate as using the original covariates on all nine studied datasets, while producing a smaller number of interpretable survival-associated clusters. The method is offered as a semi-supervised representation and clustering tool rather than as a way to improve prediction accuracy. If the claim holds, practitioners could compress large collinear covariate sets into a handful of interpretable factors without paying a predictive price.

What carries the argument

The central machinery is the weighted covariate tensor X(i,:,k) = P(X(i,:), k)^(1/2) X(i,:), where the third dimension indexes time periods (s(k-1), s(k)] and P is the Coxnet-estimated probability that the event occurs in period k. Non-negative tensor factorization then approximates this tensor by a sum of r outer products w_k ◦ h_k ◦ q_k, yielding patient meta-scores W, covariate patterns H, and temporal survival patterns Q; the meta-scores W are the final covariates for a Coxnet model. The first stage builds an unweighted tensor with indicator weights and inverse-probability-of-censoring weights to learn H and Q, and the second stage fixes H and Q to project the Coxnet-weighted tensor onto W. The survival probabilities are what make the factorization survival-aware, in contrast to plain NMF. The factorization algorithm used is Fast-HALS (fast hierarchical alternating least squares).

What would settle it

Re-run the evaluation with a strict protocol that fits the weighting Coxnet on training folds only and generates the tensor weights for test patients from that fitted model; if the c-index gap between CoxNTF and Coxnet widens substantially relative to the reported values, the comparable-accuracy claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a non-negative factorization of a survival-weighted covariate tensor produces latent meta-scores that preserve Coxnet's discriminative performance: across the nine datasets, the CoxNTF IPCW c-index is consistently close to that of Coxnet on original covariates (e.g., 0.77 vs 0.78 on flchain) and generally above CoxNMF (e.g., 0.77 vs 0.68 on flchain). The key result is that the NTF latent space carries essentially the same survival information as the original covariates, so the meta-scores can substitute for them. The paper also reports that Coxnet retains far fewer features when given NTF meta-scores than when given NMF factors (up to 45% fewer), and that the learned covariate patterns are interpretable and consistent with the hazard-ratio forest plot. The paper is explicit that CoxNTF is not meant to improve survival prediction accuracy; its purpose is joint interpretable clustering and prediction at no loss of accuracy.

Load-bearing premise

The load-bearing premise is that the Coxnet event probabilities used to weight the tensor are computed without using the survival outcomes of the patients whose predictions are later evaluated; if those probabilities see the test data, the reported match with Coxnet is optimistically biased.

Editorial extensions

If this is right

  • Coxnet trained on NTF meta-scores achieves a c-index comparable to Coxnet on original covariates across all nine datasets, so the latent representation can replace the original covariates without losing discriminative power.
  • The NTF meta-scores yield interpretable patient clusters and covariate patterns linked to early versus late events, as illustrated on the telecom churn dataset where month-to-month contracts are associated with early departure.
  • Coxnet retains substantially fewer features when given NTF meta-scores than NMF factors (up to 45% fewer), so the method compresses the covariate set.
  • CoxNTF handles redundant and collinear features without arbitrarily eliminating dummy-coded categories, avoiding the arbitrary zeroing seen in elastic-net forest plots.
  • The approach extends to time-varying covariates by modifying the weighting and tensor construction, though that extension is not tested in the paper.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The comparison with plain NMF suggests that it is the survival weighting, not the tensor structure alone, that preserves accuracy; without the Coxnet event probabilities, the factorization would reduce to an unsupervised decomposition that the paper shows performs worse.
  • The paper does not spell out that the Coxnet weighting model must be refit on training folds only to generate weights for test patients; a reader should check that the reported c-indexes came from such a clean protocol, because any leakage of test survival information into the weights would inflate the apparent match with Coxnet.
  • The sensitivity of the results to the chosen time-window percentiles and NTF rank is acknowledged in the paper but not quantified, so a natural test is whether the comparable-accuracy finding is stable across alternative choices of these hyperparameters.
  • Because the meta-scores are generic non-negative features, they could in principle be fed to any survival model, not just Coxnet, although the paper only demonstrates Coxnet.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes CoxNTF, a two-stage semi-supervised method for joint clustering and survival prediction. In the first stage, baseline covariates are placed into a three-way tensor whose time slices correspond to pre-specified survival periods; observed event times and censoring indicators determine which slice each row enters, and inverse probability of censoring weights are applied. Non-negative tensor factorization (NTF) yields covariate patterns H and temporal patterns Q. In the second stage, a Coxnet model trained on the full data supplies period-specific event probabilities used to re-weight the covariate tensor, which is then projected onto the fixed NTF bases to produce meta-scores W; a final Coxnet model is trained on W. The paper evaluates CoxNTF against Coxnet on original covariates and CoxNMF on nine datasets, reporting averaged c-indexes and numbers of retained features over 30 random partitions, and illustrates interpretable patterns on a churn dataset.

Significance. If the evaluation protocol is made outcome-free, CoxNTF would offer a modest but useful semi-supervised dimension-reduction tool: it preserves roughly Coxnet-level discrimination while providing interpretable, non-negative latent patterns, and it avoids CoxNMF's need for new-patient survival data at projection time. The authors are commendably explicit that CoxNTF is not meant to improve predictive accuracy, and they promise public code and data. However, the central quantitative claim of comparable out-of-sample prediction is not currently established because the algorithms as written use full-data survival information to build the features, and Table 1 contains no uncertainty measures. The contribution is likely salvageable with a properly nested evaluation.

major comments (3)
  1. [CoxNTF Model Training (algorithm 2) and Results] Algorithm 2, line 1 trains the Coxnet probability model on the full X, c, t, and Algorithm 1 builds the tensor whose NTF bases H and Q are returned using the full c and t. The Results section states only that each dataset was randomly partitioned into training, validation, and test sets and that criteria were averaged over 30 repetitions; it does not state that Algorithms 1 and 2 are refit on each training fold before validation/test meta-scores are computed. As written, therefore, the meta-scores W of a held-out patient are computed from H, Q, and Coxnet probabilities that have been informed by that patient's survival outcome, which would make the Table 1 c-indexes optimistically biased and would not establish out-of-sample comparability. Please specify a nested protocol in which Algorithm 1 and the Coxnet probability model are fit only on training folds, H, Q, and the probability model are frozen, and validation/test W are computed from covariates alone; then report the results under that protocol.
  2. [Results, Table 1] Table 1 reports averaged c-indexes over 30 partitions without standard deviations, confidence intervals, or paired comparisons. For example, on veterans lung cancer the NTF and COX c-indexes are 0.71 and 0.70; without an uncertainty measure this difference cannot be distinguished from sampling noise, and the claim that CoxNTF is 'always superior to CoxNMF' cannot be verified. Please report the distribution of c-indexes across the 30 repeats (e.g., mean with SD, or median with IQR) and provide paired tests for the NTF-versus-COX and NTF-versus-NMF comparisons.
  3. [Methods, NTF Rank determination and Algorithm 2] The paper does not include an ablation separating the contributions of survival weighting, tensor structure, and rank selection. Algorithm 2 weights the tensor by Coxnet probabilities before projecting onto H and Q, and the NTF rank r is chosen by validation c-index; hence the reported 'average number of features retained' is partly determined by the selected rank, and the improvement over NMF could be due to the survival weighting or to the tensor structure rather than to the full CoxNTF combination. Please add ablations such as (i) NTF on an unweighted tensor, (ii) NTF on a tensor weighted by uniform probabilities, and (iii) matching the NMF rank to the NTF rank; and report the distribution of selected ranks. This would directly support the central claim that survival-informed tensor factorization, rather than the final Coxnet step alone, drives the result.
minor comments (5)
  1. [Algorithm 1, line 13] The square-root IPCW expression is rendered with a broken equation ('× p ipcwi' and 'Taking the square root of $ipcw i$'); please fix the typesetting.
  2. [CoxNTF Model Training (algorithm 2)] The section contains an empty cross-reference ('section .'); please fill it in.
  3. [Materials] The phrases 'which observed in nearly half of the patients' appear twice with the verb 'was' missing; please correct these grammatical errors.
  4. [Results, Table 1] The table rows labeled 'NTF', 'NMF', and 'COX' should state explicitly that these are Coxnet models applied respectively to NTF meta-scores, NMF scores, and original features; the dataset size information (n, p) would be clearer in a separate column.
  5. [Discussion, Limitations] The limitation paragraph acknowledges sensitivity to the number and percentile levels of time periods but not to the NTF rank; please add rank sensitivity analyses or state the range of ranks over which the conclusions are stable.

Circularity Check

2 steps flagged · score 6.0 of 10

CoxNTF's reported comparable c-index is not shown to be out-of-sample because Algorithm 2 builds test meta-scores from Coxnet probabilities fit on the full X,c,t, making the prediction partly a refit of survival outcomes.

  1. fitted input called prediction [Methods, Algorithm 2 (CoxNTF Model Estimation), lines 1-15]
    "Input Data: The same input data as in the first stage are required ... 1: Train Coxnet survival model using X, c, t ... 4: Estimate P(X(i, :), k) using Coxnet model ... 15: Train CoxNTF model using W, c, t"

    The meta-scores W fed to the final Coxnet are built from event probabilities P(X(i,:),k) estimated by a Coxnet trained on the full X,c,t. For a held-out patient j, P(X(j,:),k) is informed by (t_j,c_j), so W_j is not outcome-free. The c-index reported in Table 1 therefore measures how well a Coxnet trained on W, which contains the first Coxnet's own in-sample survival estimates, ranks the same patients; it does not establish out-of-sample prediction. The paper never states that the Coxnet in line 1 is refit on training folds only. As written, comparability to Coxnet is partly forced by construction because the same Coxnet's risk estimates are the key input to the features.

  2. fitted input called prediction [Methods, Algorithm 1 (NTF Model Training), lines 2-16]
    "if s(k − 1) < t(i) ≤ s(k) then X (i, :, k) ← X(i, :) ... 15: Perform Nonnegative Tensor Factorization (NTF)"

    Algorithm 1 fixes the NTF basis H and Q by decomposing a tensor whose third-mode assignment is determined by observed event times t(i) and whose rows are IPCW-weighted using the Kaplan-Meier censoring distribution. Thus H and Q are functions of the full survival data (t,c). When Algorithm 2 projects any patient's covariate tensor onto these fixed H,Q, the resulting W carries survival information from all patients, including test patients, through the basis. The 'survival-consistent clustering' is therefore guaranteed by construction, and the test c-index is not an independent evaluation of a survival-free representation.

full rationale

The core methodological idea—using Coxnet probabilities as tensor weights and then NTF for clustering—is not inherently circular, and the paper honestly states that CoxNTF is not meant to improve survival prediction accuracy. If Algorithm 2's Coxnet and Algorithm 1's NTF were fit strictly on training folds, the reported c-indexes could be legitimate out-of-sample evidence. However, as written, Algorithms 1 and 2 train on X,c,t without any fold argument, and the evaluation section does not say the probability model or NTF basis are refit per fold. The central quantitative claim (that NTF meta-scores reach Coxnet-comparable c-indexes) is therefore not shown to be a survival-free prediction; it can be explained by the test patients' own outcomes entering the feature construction through P and through H,Q. The limitations paragraph addresses time-window and rank sensitivity but not this outcome-free requirement. There are no load-bearing self-citations or imported uniqueness theorems here; the circularity is of the fitted-input/prediction-leakage type, and it affects the paper's headline comparison.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

CoxNTF rests on two fitted hyperparameters (rank and period boundaries), an unspecified Coxnet regularization setting, and standard NTF, Cox, and IPCW assumptions. The most consequential unverified premise is the clean cross-validation of the probability-weighting step.

free parameters (3)
  • NTF rank r = selected per dataset by validation c-index, not tabulated
    The number of components is a data-driven hyperparameter; comparisons are therefore tuned on validation, and the final c-index is not corrected for this selection.
  • survival period boundaries s = 10th, 25th, 50th, 75th, 90th percentiles of observed times
    Chosen to balance sample sizes across periods; the paper lists the impact of this choice as under-explored.
  • Coxnet regularization parameters (alpha, lambda) = not reported
    Used both to estimate event probabilities and to fit the final model; without fixed values the protocol is not fully reproducible.
assumptions (4)
  • domain assumption The three-way tensor is well approximated by a rank-r CP decomposition with non-negative factors
    Invoked in Algorithm 1, steps 15-16; approximation quality is not reported.
  • domain assumption Censoring is independent and Kaplan-Meier IPCW weights correct for it when applied as sqrt(ipcw)
    Used in Algorithm 1, steps 11-13; standard but unverified for each dataset.
  • domain assumption Cox proportional hazards and elastic-net shrinkage apply to all datasets
    Used for probability estimates and the final CoxNTF model; no proportional-hazards diagnostics are reported.
  • ad hoc to paper Event time can be discretized into fixed percentile windows without loss of relevant survival information
    The Methods section defines s via 10/25/50/75/90 percentiles; no sensitivity analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoxNTF: A New Approach for Joint Clustering and Prediction in Survival Analysis." pith.science (2026). https://pith.science/paper/MTPQTUXP

@misc{pith2026250606411,
  author       = {Pith},
  title        = {Pith review of: CoxNTF: A New Approach for Joint Clustering and Prediction in Survival Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MTPQTUXP}},
  note         = {Machine review of arXiv:2506.06411}
}
read the original abstract

The interpretation of the results of survival analysis often benefits from latent factor representations of baseline covariates. However, existing methods, such as Nonnegative Matrix Factorization (NMF), do not incorporate survival information, limiting their predictive power. We present CoxNTF, a novel approach that uses non-negative tensor factorization (NTF) to derive meaningful latent representations that are closely associated with survival outcomes. CoxNTF constructs a weighted covariate tensor in which survival probabilities derived from the Coxnet model are used to guide the tensorization process. Our results show that CoxNTF achieves survival prediction performance comparable to using Coxnet with the original covariates, while providing a structured and interpretable clustering framework. In addition, the new approach effectively handles feature redundancy, making it a powerful tool for joint clustering and prediction in survival analysis.

Figures

Figures reproduced from arXiv: 2506.06411 by the authors.

Figure 1
Figure 1. Interdependence between covariate matrix, tensor [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. NTF patterns found in the ”ds2” analysis [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Coxnet forest plot The Coxnet forest plot of hazard ratios is shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 8 canonical work pages

  1. [1]

    R.; and Mesirov, J

    Brunet, J.-P.; Tamayo, P.; Golub, T. R.; and Mesirov, J. P. 2004. Metagenes and molecular pattern discovery using matrix factorization. Proceedings of the National Academy of Sciences 101(12): 4164--4169. doi:10.1073/pnas.0308531101

    doi:10.1073/pnas.0308531101 2004

  2. [2]

    Cai, D.; He, X.; Han, J.; and Huang, T. S. 2011. Graph Regularized Nonnegative Matrix Factorization for Data Representation. IEEE Transactions on Pattern Analysis and Machine Intelligence 33(8): 1548--1560. doi:10.1109/TPAMI.2010.231

    doi:10.1109/tpami.2010.231 2011

  3. [3]

    Cichocki, A.; and Phan, A.-H. 2009. Fast Local Algorithms for Large Scale Nonnegative Matrix and Tensor Factorizations. IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences E92.A(3): 708--721. doi:10.1587/transfun.E92.A.708

    doi:10.1587/transfun.e92.a.708 2009

  4. [4]

    S.; Bergh, J.; Lidereau, R.; Ellis, P

    Desmedt, C.; Piette, F.; Loi, S.; Wang, Y.; Lallemand, F.; Haibe-Kains, B.; Viale, G.; Delorenzi, M.; Zhang, Y.; D'assignies, M. S.; Bergh, J.; Lidereau, R.; Ellis, P. A.; Harris, A. L.; Klijn, J. G. M.; Foekens, J. A.; Cardoso, F.; Piccart, M. J.; Buyse, M. E.; and Sotiriou, C. 2007. Strong Time Dependence of the 76-Gene Prognostic Signature for Node-Neg...

    work page 2007

  5. [5]

    H.; Li, T.; and Jordan, M

    Ding, C. H.; Li, T.; and Jordan, M. I. 2010. Convex and Semi-Nonnegative Matrix Factorizations. IEEE Transactions on Pattern Analysis and Machine Intelligence 32(1): 45--55. doi:10.1109/TPAMI.2008.277

    doi:10.1109/tpami.2008.277 2010

  6. [6]

    A.; Kyle, R

    Dispenzieri, A.; Katzmann, J. A.; Kyle, R. A.; Larson, D. R.; Therneau, T. M.; Colby, C. L.; Clark, R. J.; Mead, G. P.; Kumar, S. K.; Melton, L. J.; and Rajkumar, S. V. 2012. Use of nonclonal serum immunoglobulin free light chains to predict overall survival in the general population. Mayo Clinic proceedings 87 6: 517--23. ://pubmed.ncbi.nlm.nih.gov/22677072/

    work page arXiv 2012

  7. [7]

    E.; Lee, K

    Harrell, F. E.; Lee, K. L.; and Mark, D. B. 1996. Multivariable prognostic models: Issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Statistics in Medicine 15(4): 361--387. doi:10.1002/(SICI)1097-0258(19960229)15:4<361::AID-SIM168>3.0.CO;2-4

    doi:10.1002/(sici)1097-0258(19960229)15:4 1996

  8. [8]

    W.; Lemeshow, S.; and May, S

    Hosmer, D. W.; Lemeshow, S.; and May, S. J. 2008. Applied Survival Analysis: Regression Modeling of Time-to-Event Data. ://academic.oup.com/biometrics/article-abstract/65/2/671/7331827

    work page 2008

Show all 18 references
  1. [9]

    Hou, L.; Chu, D.; and Liao, L.-Z. 2024. Convergence of a Fast Hierarchical Alternating Least Squares Algorithm for Nonnegative Matrix Factorization. IEEE Transactions on Knowledge and Data Engineering 36(1): 77--89. doi:10.1109/TKDE.2023.3279369

    2024

  2. [10]

    Hoyer, P. O. 2004. Non-negative Matrix Factorization with Sparseness Constraints. Journal of Machine Learning Research 5: 1457--1469. ://jmlr.org/papers/v5/hoyer04a.html

    2004

  3. [11]

    Huang, Z.; Salama, P.; Shao, W.; Zhang, J.; and Huang, K. 2020. Low-Rank Reorganization via Proportional Hazards Non-negative Matrix Factorization Unveils Survival Associated Gene Clusters. ://arxiv.org/abs/2008.03776

    2020 arXiv

  4. [12]

    D.; and Prentice, R

    Kalbfleisch, J. D.; and Prentice, R. L. 2002. The Statistical Analysis of Failure Time Data: Kalbfleisch/The Statistical. ://onlinelibrary.wiley.com/doi/book/10.1002/9781118032985

    2002 doi

  5. [13]

    D.; and Seung, H

    Lee, D. D.; and Seung, H. S. 1999. Learning the parts of objects by non-negative matrix factorization. Nature 401(6755): 788--791. doi:10.1038/44565

    1999 doi

  6. [14]

    Li, Z.; Meng, Z.; Xiao, L.; Du, J.; Jiang, D.; and Liu, B. 2024. Constructing and identifying an eighteen-gene tumor microenvironment prognostic model for non-small cell lung cancer. World Journal of Surgical Oncology 22: 319. doi:10.1186/s12957-024-03588-y

    2024 doi

  7. [15]

    Pascual-Montano, A.; Carazo, J.; Kochi, K.; Lehmann, D.; and Pascual-Marqui, R. 2006. Nonsmooth nonnegative matrix factorization (nsNMF). IEEE Transactions on Pattern Analysis and Machine Intelligence 28(3): 403--415. doi:10.1109/TPAMI.2006.60

    2006 doi

  8. [16]

    Schumacher, M.; Bastert, G.; Bojar, H.; H \"u bner, K.; Olschewski, M.; Sauerbrei, W.; Schmoor, C.; Beyerle, C.; Neumann, R. L. A.; and Hf, R. 1994. Randomized 2 x 2 trial evaluating hormonal treatment and the duration of chemotherapy in node-positive breast cancer patients. G...

    1994

  9. [17]

    Simon, N.; Friedman, J.; Hastie, T.; and Tibshirani, R. 2011. Regularization Paths for Cox's Proportional Hazards Model via Coordinate Descent. Journal of Statistical Software 39(5): 1--13. doi:10.18637/jss.v039.i05

    2011 doi

  10. [18]

    J.; D’Agostino, R

    Uno, H.; Cai, T.; Pencina, M. J.; D’Agostino, R. B.; and Wei, L.-J. 2011. On the C‐statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Statistics in Medicine 30. ://onlinelibrary.wiley.com/doi/10.1002/sim.4154

    2011 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.