acceptodds
Under review as a conference paper at ICLR 2027

CalibCF: Graph-Conditioned Spectral Cutoffs for Training-Free Recommendation

Abstract

Training-free spectral recommenders can rival learned embedding models, yet their accuracy depends on where the spectrum of the user-item graph is cut. Across five benchmarks, the best retained dimension spans 114–1,869, making a fixed rank difficult to transfer between graphs. The upper edge of the noise spectrum offers a graph-specific reference for this decision, but standard estimators are poorly suited to recommendation: user and item degrees create heterogeneous noise, while scalable methods compute only a short spectral head rather than the full bulk. We introduce CalibCF, which estimates the noise edge from the interaction graph and the head already available to a spectral recommender. User and item degrees define a separable variance profile, from which CalibCF derives a graph-conditioned template for the noise spectrum. After excluding leading signal directions, a one-parameter regression fits the template's scale to the observed squared singular values. We prove that the resulting edge estimate is consistent under a degree-corrected Bernoulli graph model and spectral regularity conditions. CalibCF then sets the operating cutoff relative to this edge and applies a continuous, closed-form shrinkage response to the retained directions. Using spectral heads covering only 4.9–10.4% of the spectrum, the template predicts the uncomputed eigenvalue distribution on four benchmarks with KS distance 0.016–0.049, a 3–12× improvement over an identically fitted Marchenko–Pastur template. CalibCF outperforms all seven compared recommenders on five benchmarks, improving NDCG@20 over the strongest baseline by up to 4.0%; 65 of 66 paired comparisons are significant after Holm correction. With a fixed operating point and no target-domain validation, it retains 98.3% of its per-graph tuned NDCG@20, compared with 96.7% for the best fixed absolute rank.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Collaborative Filtering, Spectral Methods, Training-Free Recommendation, Random Matrix Theory

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_MH7qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CalibCF is a training-free spectral collaborative filtering method. The central idea is to let the user/item degrees tell you the shape of the noise spectrum, then use the singular values already computed in a short spectral head to calibrate its scale. The method uses that estimated noise edge to set a cutoff and apply a continuous shrinkage rule.

The paper backs this up with theoretical results under a degree-corrected random-graph model, and experiments on five datasets, including full-catalog recommendation metrics, baseline comparisons, ablations, transfer tests, and spectral-fit analyses. From the code I reviewed, the main calibration and scoring formulas appear to match the paper.

Critical strengths
  1. I found the main idea clean and useful. Separating noise-spectrum shape from scale is a nice way to make spectral calibration depend on the graph, rather than relying on a rank or cutoff that may not transfer across datasets. The method also targets a practical setting: only a truncated spectral head is available.

  2. The experiments test more than just the final recommendation score. Alongside five datasets and several strong baselines, the paper studies spectral fit, ablations, head size, and transfer. That makes the evaluation feel meaningfully tied to the method’s central claims. The results look promising overall, though I would not describe every win as decisive or statistically significant.

PS. Code is available.

Critical weaknesses
  1. The evaluation protocol and the supplied code do not line up cleanly, and the full results are hard to reproduce from this ZIP. The main text describes a per-user 70/10/20 split, while the code uses the existing public test split for Gowalla and Yelp and holds validation data out of the training split. The appendix also refers to public splits, so this may be a reporting inconsistency rather than a flawed experiment—but it needs to be made explicit. The ZIP also lacks the baseline, significance-testing, and appendix-analysis scripts, so I could not independently reproduce the paper’s full set of comparisons.

  2. Some of the headline evidence is less conclusive than the broad claims suggest. On ML-10M, CalibCF’s NDCG@20 is essentially tied with EASE and the reported difference is not statistically significant. More broadly, the theoretical guarantee rests on graph and spectral assumptions that may not hold for very sparse, heavy-tailed recommendation graphs, and the paper offers limited evidence about how well those assumptions describe its real datasets.

Questions for authors
  1. Could you clarify the exact train/validation/test protocol used for each dataset—especially Gowalla and Yelp—and reconcile the main text’s 70/10/20 description with the public splits used in the code?

  2. Could you provide the scripts and exact configurations used to generate the baseline results, significance analyses, and appendix experiments? In particular, what SVD implementations and settings produced the runtime numbers?

  3. How well do the degree-conditioned template and edge estimate hold up on real, very sparse or heavy-tailed graphs? Do you have diagnostics that would let a practitioner tell when the separable-template assumptions are a poor fit?

  4. What is the motivation for the fixed 1.02 margin in the exclusion update, and does the calibration or recommendation quality change materially for nearby values?

  5. The paper calls CalibCF “training-free,” but it still computes a truncated SVD and selects \kappa and \gamma on a validation split. Could you clarify what “training-free” means here—specifically, whether it means no gradient-based parameter learning—and help isolate how much of the gain comes from the degree-conditioned edge calibration rather than the low-rank SVD itself?

Reviewer_FyVAOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

CalibCF is a training-free spectral recommender that calibrates its cutoff using the interaction graph and a computed spectral head. It constructs a degree-informed, separable variance template and fits its scale to observed squared singular values after excluding leading signal directions. The estimated noise edge then anchors a relative cutoff and a continuous singular-value shrinkage function.

The paper provides conditional consistency and score-stability guarantees, alongside experiments on five recommendation benchmarks, unseen-spectrum prediction, component controls, and cross-dataset transfer. Edge calibration itself does not require validation labels; the paper separately evaluates validation-tuned and fixed operating points.

Critical strengths
  1. The method addresses two concrete obstacles: heterogeneous interaction noise and access to only a truncated spectrum. Combining graph-derived variance information with a one-parameter spectral fit provides a principled reference for cutoff selection without reconstructing the missing empirical tail. This is a meaningful contribution beyond introducing another spectral response function.

  2. The unseen-spectrum and controlled synthetic experiments directly investigate the proposed mechanism, rather than relying exclusively on ranking improvements. The head-width analysis and transfer results further support its practical value: the calibrated coordinate is relatively stable and, in the reported comparisons, transfers better than absolute rank. Together, these results make the calibration hypothesis credible.

Critical weaknesses
  1. The issue is not that noise has a distribution, but whether the assumed distribution adequately distinguishes stochastic fluctuations from weak, useful preference structure. The theory relies on independent Bernoulli interactions, fixed-rank preferences, approximately separable variances, and additional spectral regularity. The reported synthetic evaluation primarily varies degree heterogeneity, while matching the observed real-data spectrum does not directly validate the latent noise decomposition. It therefore remains unclear how reliably the calibration behaves when numerous weak preference directions contaminate its fitting window.

  2. Although the paper includes substantial controls, the MP-template comparison primarily evaluates spectral prediction rather than recommendation and transfer under an otherwise identical pipeline. Better spectral fit does not necessarily imply better ranking. Additionally, the LODO absolute-rank search is capped at 512, leaving uncertainty about how much this restriction contributes to the reported transfer advantage. These gaps limit the attribution of downstream gains specifically to the proposed calibration.

Questions for authors
  1. Do any experiments explicitly violate the fixed-rank or approximately separable-variance assumptions, rather than varying degree heterogeneity alone? In particular, how do edge-estimation error and recommendation accuracy change when the graph contains many weak but predictive preference directions? This would clarify the practical scope of the calibration guarantee.

  2. How does an identically fitted MP template perform when the operator, spectral head, shrinkage function, and readout are held fixed, particularly under fixed and LODO operating-point selection? Does the transfer advantage also persist when the absolute-rank search extends beyond 512 using sufficiently large, matched spectral heads? These comparisons would materially affect my assessment of the core contribution.

2 comments
Area_Chair_q1UyOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer FyVA,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Thank you. This addressed all of my questions. I would like to increase my score!

Reviewer_e9mkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CalibCF, a training-free spectral recommendation method that automatically determines appropriate spectral cutoffs for different user-item interaction graphs. The key idea is to construct a graph-conditioned noise spectral template using user and item degree distributions and estimate the noise edge from a small number of leading singular values, without computing the full spectrum. Experiments on five benchmark datasets demonstrate that CalibCF outperforms several training-free and learned recommendation baselines while achieving accurate spectral estimation, efficient computation, and relatively stable performance across different graphs.

Critical strengths
  1. The problem is clearly defined, the technical approach is well motivated. Rather than simply introducing a new spectral filtering function, the paper addresses a key challenge in spectral recommendation: how to determine an appropriate spectral cutoff across different interaction graphs.
  2. The paper not only evaluates recommendation performance on five datasets but also validates the proposed method from multiple perspectives. In particular, the authors demonstrate that a stable estimate of the noise edge can be obtained using only a small portion of the spectrum, avoiding the need to compute the full spectrum.
Critical weaknesses
  1. The main innovation of the paper lies in constructing a noise spectral template from the graph's degree distribution, enabling more accurate estimation of the noise spectrum. Although Figure 2 demonstrates that the degree-based template more accurately predicts the uncomputed spectrum, the recommendation ablation studies mainly compare CalibCF with the spectral edge estimator of eOptShrink and hard truncation. The contribution of the graph-conditioned template to recommendation performance is therefore not directly isolated.
  2. The paper emphasizes that CalibCF can use a shared κ\kappa across different datasets, avoiding dataset-specific tuning of the truncation rank. However, the results on ML-10M show a noticeable deviation from this central claim, raising concerns about the generalizability of the proposed calibration strategy across different interaction graphs.
Questions for authors

If the estimated noise edge represents the boundary between signal and noise, why do the optimal cutoff multipliers consistently fall below 1? In particular, why does ML-10M require κ=0.7\kappa=0.7, while the other datasets favor values around 0.85–0.9? Can the authors explain this difference and propose a label-free way to select κ\kappa?

4 comments

Thank you for the additional experiments. The new results provide useful evidence supporting the cross-dataset transferability of the shared multiplier, and they substantially address my original concern.

However, my second concern remains insufficiently addressed. Specifically, the response explains why the optimal multiplier may fall below 1, but does not explain why ML-10M exhibits a noticeably different optimal multiplier from the other datasets. Besides, while the average transferred performance is encouraging, the response does not report the dataset-specific performance retention under the shared multiplier. Moreover, the extended rank comparison shows that the transferred absolute rank outperforms the transferred multiplier on ML-10M (0.3515 vs. 0.3351), suggesting that the proposed calibration strategy may be less effective on this particular graph.

Thank you for the follow-up and for your assessment of the new experiments. We address your three remaining points in the order you raised them; the first one prompted an additional experiment.

Why ML-10M differs from the other graphs. Thank you for pressing on this point. To answer it directly, we repeated the band experiment of Appendix F.7 on ML-1M, the other MovieLens graph, providing a same-domain comparison under matched preprocessing. The rule is exactly the same: the weights of κ = 0.7 are held fixed and the directions below 0.9 b̂ are added band by band, on three splits (retained dimension in parentheses).

ML-1MSplit 1 (main)Split 2Split 3
Directions above 0.9 b̂, weights of κ = 0.9 (fixed and selected point)0.3471 (114)0.3398 (111)0.3414 (115)
Same directions, weights of κ = 0.70.34290.33550.3373
+ directions in (0.8 b̂, 0.9 b̂]0.3318 (265)0.3275 (258)0.3283 (264)
+ directions in (0.7 b̂, 0.8 b̂]0.3266 (498)0.3231 (491)0.3241 (499)

On ML-1M each added band lowers NDCG@20, by 0.0042–0.0111; on ML-10M each of the same two bands raises it, by 0.0071–0.0083 (table in our previous reply). The bands contain a comparable number of directions on the two graphs (384 and 492 on the main splits). These opposite effects provide an empirical explanation for the different validation-selected multipliers: under the tested response, ML-10M benefits from retaining a deeper band below the edge, whereas ML-1M favors a cut closer to the calibrated edge. This is consistent with the mechanism illustrated by the stress test in Appendix H.6, where the best multiplier moves below the edge once many weak preference directions lie under it.

Per-dataset retention. Thank you for asking for the dataset-level numbers. For completeness, we distinguish the fixed shared default (κ, γ) = (0.9, 0.5) from the leave-one-dataset-out (LODO) operating point of Table 3 and report both, relative to the value selected on each graph's own validation set (Table 2):

DatasetSelected on validationFixed defaultRetentionLODORetention
ML-1M0.34710.3471100.0%0.3488100.5%
Gowalla0.16240.162299.9%0.149091.7%
Yelp20180.06200.060297.1%0.060397.3%
ML-10M0.35550.336994.8%0.335194.3%
Pinterest0.12300.122799.8%0.121598.8%
Mean98.3%96.5%

The fixed default retains 97.1–100.0% on four graphs and 94.8% on ML-10M, the graph discussed above (a value above 100% can occur because the reference is selected on validation, not on test). On Gowalla and ML-10M the LODO point has the same cutoff as the fixed default (κ = 0.9) and differs only in the readout (γ = 0.25 instead of 0.5), so the lower Gowalla value comes from the transferred readout and not from the multiplier. With the readout held at its default (γ = 0.5), the LODO multiplier is κ = 0.85 for every target graph and retains 100.5%, 100.0%, 97.3%, 96.6% and 98.8% in the order of the table (mean 98.6%; Appendix G.1; Appendix H.7, Table 35). This readout-fixed comparison isolates cutoff transfer; it does not replace the joint-LODO results reported above.

The transferred rank on ML-10M. You are right that on this graph the transferred rank scores above the transferred multiplier (0.3515 against 0.3351). This result is consistent with the band-level findings above. On ML-10M, the LODO-selected rank of the extended grid retains 1,024 directions, whereas the LODO-selected multiplier setting retains 117. The former includes the bands below the edge shown above to improve ranking on this graph, so the result is in line with the benefit of a deeper cut there. The comparison establishes an advantage for the transferred rank setting on ML-10M; it does not by itself establish an error in the calibrated noise edge, and the held-out spectral diagnostic separately supports the fit of the degree-conditioned template on this graph (residual 1.1%, against 13.3% for an MP template; Appendix H.4). The advantage does not carry over to graphs with a different best dimension: the same transferred rank of 1,024 retains 80.1% on ML-1M, whose selected dimension is 114. We therefore present the shared multiplier as the label-free default, and the per-graph selection of Table 2 as the way to reach the additional structure of a graph such as ML-10M when a validation set is available.

Following your question, we will add the ML-1M decomposition to Table 16 (Appendix F.7) and update the revised PDF accordingly. Thank you again for the careful reading, which helped us make this part of the paper clearer.

Thank you very much for your prompt and thoughtful response! I am genuinely delighted to see such a niche yet fascinating research direction in recommender systems being explored in the era of large language models. I have learned a lot from your paper, particularly about concepts and techniques I was previously unfamiliar with.

Your response has addressed most of my concerns, and I have no further questions at this stage. Therefore, I have decided to update my final recommendation to Weak Accept.

Thank you again for your efforts, and best of luck with your paper!

Thank you very much for your kind words. We sincerely appreciate the time and care you put into reviewing our paper and into the discussion. Your comments have helped us improve the work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.