acceptodds
Under review as a conference paper at ICLR 2027

Readable Factors, Unresolved Combinations: Support-Aware Correction for Warm-Start Recommendation

Abstract

A warm item can combine factors a user has seen separately but never together, even when that combination occurs elsewhere in training. Standard warm-start metrics mix these cases with repeated combinations, while tests of globally unseen combinations study a different setting. We define a pre-cutoff evaluation cohort and use an independently fitted audit to test whether pair features add held-out ranking value beyond a frozen ranker and the individual factors. On H&M fashion and M3L movies, the audit finds a positive increment from purified pair features; pair permutations and equal-capacity random controls are near zero. We introduce Support-Aware Compositional Factor Residual (SCFR), which removes declared lower-order variation from pair features and uses train-only pairwise-logistic curvature to shrink weakly supported comparison directions. One correction scores every eligible warm-catalog item. The analysis states when a target comparison is identifiable and locally stable and gives a conditional squared-risk result for shrinkage when signal aligns with support. On the two confirmatory anchors, SCFR improves target NDCG@20 over the strongest matched-input control selected on validation by +0.0074 [0.0015, 0.0081] and +0.0083 [0.0014, 0.0094], corresponding to relative gains of 20.1% and 6.7%, while passing prespecified overall and non-target non-inferiority checks. Amazon Sports replicates the target-gain direction; controlled anti-alignment and weak-support tests show where the advantage vanishes or reverses.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
recommender systems, compositional generalization, model auditing, interaction modeling, regularized estimation

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_KHx5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies warm-start recommendation for combinations that are globally observed but new to a particular user. SCFR adds a learned correction to a frozen ranker, combining lower-order residualization of factor-pair features with curvature-based directional regularization. The experiments report target-cohort improvements alongside overall and non-target non-inferiority checks.

Critical strengths
  • Separating redundant lower-order information before regularizing weakly supported directions is a coherent approach. Using comparison-loss curvature to guide regularization provides an interesting statistical motivation.

  • The supplementary experiments are detailed, with extensive experimental analyses and descriptions of the experimental conditions.

Critical weaknesses
  1. Clarify the correction setting and its objective. I may be missing some context in recommendation, but the intended application and objective of the proposed correction remain unclear to me. (i) If it is intended for fine-ranking, freezing an existing ranker and adding a correction to its prediction scores does not seem to me to be a typical formulation. For comparison, PACC models position bias within a jointly trained click/conversion predictor [1], rather than freezing a ranker and adding a learned correction term based on specific factor combinations to its prediction scores. (ii) If it is intended for calibration in applications such as advertising, applying a shared strictly increasing calibration function preserves the relative ordering of prediction scores. How does the proposed correction relate to this setting and to calibration methods such as AdaCalib and DESC [2,3]? (iii) If it is intended to support feature engineering, post-hoc feature attribution generally analyzes model predictions rather than directly correcting recommendation outputs [4]. I therefore encourage the authors to provide more related work explaining when freezing the original ranker and adding a correction term to its scores is appropriate, and to briefly introduce the task background in the Abstract or Introduction.

  2. Review the writing and logical flow more carefully. The prose has noticeable features of AI-assisted writing, and some passages interrupt the flow of the argument or make abrupt logical transitions. For example, the manuscript repeatedly emphasizes that its theory does not guarantee NDCG improvement, giving parts of the discussion a repetitive, defensive tone. The authors should review and revise these passages more carefully.

  3. Validate on larger datasets. The reported numbers of eligible warm users and items in each domain are below 100,000 (Table 6), which is a small scale for recommendation scenarios. I recommend validating the results on a larger dataset, such as AliCCP.

  4. Include more feature-interaction baselines. The baseline comparisons are insufficient. Since the paper focuses on feature combinations, established feature-interaction models such as Wide & Deep [5] and FinalMLP [6] should be considered. The paper includes DCNv2-head as a correction module, but how does the original DCN-V2 perform? Please include this comparison or provide references supporting the rationale for the current baseline selection.

References

[1] Yibo Wang et al. Click-Conversion Multi-Task Model with Position Bias Mitigation for Sponsored Search in eCommerce. SIGIR, 2023.

[2] Penghui Wei et al. Posterior Probability Matters: Doubly-Adaptive Calibration for Neural Predictions in Online Advertising. SIGIR, 2022.

[3] Shuai Yang et al. Deep Ensemble Shape Calibration: Multi-Field Post-hoc Calibration in Online Advertising. KDD, 2024.

[4] Ian Covert, Scott Lundberg, and Su-In Lee. Explaining by Removing: A Unified Framework for Model Explanation. JMLR, 22(209):1–90, 2021.

[5] Heng-Tze Cheng et al. Wide & Deep Learning for Recommender Systems. 2016. Paper.

[6] Kelong Mao et al. FinalMLP: An Enhanced Two-Stream MLP Model for CTR Prediction. AAAI, 2023.(https://ojs.aaai.org/index.php/AAAI/article/view/25577).

Questions for authors

See the weakness.

Reviewer_vnawOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Support-Aware Compositional Factor Residual (SCFR), a ranking correction method for warm-start recommendation. The method aims to address user-local unseen factor combinations, where a user has seen individual item factors but not their combination. SCFR removes lower-order variation from pair features and uses pairwise-logistic curvature to shrink weakly supported comparison directions. Experiments on H&M and M3L, with Amazon Sports as an external replication, show improvements in target NDCG@20 over matched-input controls while satisfying overall and non-target non-inferiority checks.

Critical strengths
  1. The paper introduces a user-local composition cohort and uses an independent audit to examine whether factor pairs provide additional ranking value beyond individual factors.
  2. The paper evaluates SCFR against matched-input controls and includes ablation studies, robustness tests, and non-inferiority checks for overall and non-target recommendation performance.
Critical weaknesses
  1. Insufficient motivation and unclear advantages over existing residual-based recommendation methods. The motivation of using residual information to capture what existing representations cannot explain is not new. Previous recommendation methods have already explored residual modeling, such as SID-based generative recommendation methods (e.g., TIGER and LETTER), which use hierarchical residual quantization to learn structured item representations and improve generalization. Although these methods use residuals differently from SCFR, they share the general motivation of capturing additional information through residual modeling. However, the paper does not sufficiently discuss these related methods or explain why its specific residual correction is more suitable for user-local unseen factor combinations. More importantly, the advantages of SCFR over existing residual-based recommendation approaches remain unclear. This weakens the motivation for the proposed method and makes its contribution relative to existing work difficult to assess.
  2. Limited generality and unclear practical significance of the proposed setting. The proposed setting relies on predefined factor combinations, such as product type × color for H&M and genre pairs for M3L. However, real user preferences may involve more complex factors and combinations that cannot be covered by these manually selected schemas. The method does not explain how such factors can be automatically identified or how it generalizes beyond the predefined combinations. Therefore, it remains unclear whether the observed improvements specifically address user-local unseen combinations or reflect a more general benefit from factor interaction modeling. These limitations weaken the practical significance of the proposed problem and method.
Questions for authors
  1. How does SCFR compare with existing SID-based recommendation methods, such as TIGER and LETTER, on the proposed user-local unseen combination setting?
  2. How can SCFR identify and model useful factor combinations that are not covered by the manually predefined factor schemas?
Reviewer_KBiiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a class of samples in warm-start recommendation: every attribute of an item has appeared separately in the user's history, yet the combination of these attributes is new to that user (the combination has appeared for other users). The authors propose Support-Aware Compositional Factor Residual (SCFR): it first separates attribute main effects from compositional effects, and then uses comparison information from the training data to impose constraints of different strength on different directions, so as to improve the ranking of such samples. Experiments are conducted on H&M, M3L and Amazon Sports, and the method is validated through theoretical analysis, ablation studies and robustness tests.

Critical strengths
  1. Clear problem definition. The paper separates "attribute combinations absent from the global training data" from "combinations that are merely new to a given user", and focuses on the latter: the user, the item, all attributes and the combination are already warm globally, and only this combination is new to that user. This setting gives a finer window onto the ranking behaviour that aggregate warm-start metrics obscure.

  2. The method is grounded in theoretical analysis. SCFR first residualizes the combination features against the declared lower-order attribute span, and then, based on the loss curvature of the pairwise comparisons in the training data, applies shrinkage of different strength to different directions. The paper further analyses identifiability, local stability, and the alignment between signal and support.

  3. The experimental design is systematic. The authors retrain existing recommenders under the same protocol as context, and use matched-input controls that share the base score, the factor interface, the sampler and the tuning budget as the primary comparison. They additionally conduct an independent audit, ablation studies, replications across cutoffs and with a second backbone, and failure-case analyses. Overall and non-target performance outside the target samples is also reported.

Critical weaknesses
  1. The scope of the method is still limited. The target samples are obtained through fairly strict filtering and make up only a small share of the evaluation positives, and the gains of SCFR are concentrated on these samples. The paper does report how the gain varies with the coverage of the target samples, but the pattern is not stable: the largest gain occurs at the coverage used in the main experiment, while nearby settings are markedly weaker and in places no longer significant. What needs to be clarified is whether the gain holds as coverage changes, and whether it remains measurable outside the target samples.

  2. The results are quite sensitive to how the attribute combinations and the support condition are defined. The sensitivity analysis in the appendix shows that when the evaluated item is removed from the support evidence, the gain on one of the datasets drops sharply and is no longer significant, and that coarser-grained or semantically similar attribute definitions make some results insignificant. This raises the question of whether the advantage comes from transferable compositional regularities, or depends on a particular attribute partition and on how the samples are filtered. The authors are encouraged to explain why the main experiment adopts a support definition that includes the evaluated item itself.

  3. There remains a gap between the theoretical analysis and the actual optimization objective. The conclusion that directional shrinkage reduces predictive risk rests on fixed-design, squared-loss and specific noise and covariance assumptions, whereas the actual task uses a ranking objective and NDCG. The authors have stated this boundary, and the synthetic intervention does examine the mechanism, but on real data only the association between the proposed diagnostics and later-cutoff error is examined, while the alignment condition behind the theoretical result is not itself estimated. In addition, the ablation against the unweighted Gram matrix also changes the scale of the matrix, so it cannot fully isolate the contribution of the curvature information itself; a scale-matched control would help.

  4. Writing and reproducibility can still be improved. The main text is dense in terminology, mathematical notation and experimental protocols, some key implementation details are scattered across a fairly long appendix, and the main text and the appendix partly overlap, which adds to the difficulty of reading and reproducing the work. We suggest simplifying the description of the method, stating the key variables and training steps in the main text, and merging the duplicated tables. Moreover, no experimental code is provided, which limits independent verification of the results.

Questions for authors

See Weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.