acceptodds
Under review as a conference paper at ICLR 2027

Null-Space Calibration of Auxiliary Evidence for Multi-Behavior Recommendation

Abstract

Multi-behavior recommendation learns user and item representations from target and auxiliary interactions to improve target prediction. Auxiliary representations may encode preferences not fully reflected in target interactions, but association strength alone does not determine the ranking value of this evidence. A key challenge is to retain target-associated information while learning to use potentially complementary evidence. To this end, we propose Target-referenced Null-space Calibration (TNC), a plug-in module that calibrates auxiliary representations through null-space constraints. Using cross-covariance between paired auxiliary and target representations, TNC constructs a target-associated subspace through truncated singular value decomposition. It retains both the target-associated and null-space components and learns separate calibration gains under target ranking supervision. For a fixed representation, reference, and target-associated gain, adjusting the null-space gain calibrates potentially complementary evidence without changing the prescribed target-associated projection. The module operates between behavior encoding and fusion and is jointly trained with compatible encoders, allowing target supervision to guide both representation learning and calibration. Experiments on three real-world datasets show that integrating TNC improves ranking performance across multiple multi-behavior recommendation backbones. To facilitate reproducibility, our implementation is available in an anonymous repository at this url.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-behavior Recommendation, Null-space Projection, Representation Calibration, Cross-behavior Knowledge Transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_M7jMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers auxiliary representation calibration for multi-behavior recommendation. The main motivation is that information strongly associated with the target behavior may reinforce existing preferences, while less-associated information may still contain complementary signals useful for ranking. The paper proposes TNC, which constructs a target-associated subspace from auxiliary–target cross-covariance using truncated SVD, and decomposes auxiliary representations into target-associated and null-space components. Separate learnable gains are then used to calibrate the two components under target-ranking supervision. TNC is designed as a plug-in module between behavior encoding and fusion and can be jointly trained with different multi-behavior recommendation backbones. Experiments on Taobao, Tmall, and Beibei with six backbone models show improvements in most settings, with additional ablations examining the main components of the proposed method.

Critical strengths
  1. The paper clearly distinguishes target association from ranking value, and the proposed TNC framework is well aligned with this motivation.
  2. TNC is evaluated on multiple backbones and real-world datasets, with generally consistent improvements and useful ablation studies.
  3. The null-space formulation provides an intuitive explanation of the calibration mechanism, and the theoretical analysis further supports the proposed design.
Critical weaknesses
  1. The null-space component is geometrically orthogonal to the selected target-associated subspace, but orthogonality itself does not necessarily imply that it contains complementary preference information. Although the ablation results show that retaining this component improves performance, more direct evidence would help verify that it indeed captures complementary signals rather than general residual information or noise.
  2. Some integrations also replace the original aggregation or scoring functions, which may confound the comparison.
  3. TNC constructs the target-associated subspace using linear cross-covariance and truncated SVD, and the formulation assumes compatible feature coordinates between user and item representations within each behavior. This may limit its ability to capture nonlinear auxiliary–target relationships and could restrict its applicability to backbones with heterogeneous representation spaces.
Questions for authors
  1. Could the null-space component contain redundant information or noise in addition to complementary preference signals, and if so, how would they affect the calibration and ranking performance?
  2. To what extent are the reported improvements attributable to TNC itself rather than the modified aggregation or scoring functions?
  3. How sensitive is TNC to the linear cross-covariance and truncated SVD used for target-reference construction? Would nonlinear auxiliary–target dependencies affect the quality of the constructed reference subspace?
Reviewer_eCFVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Target-referenced Null-space Calibration (TNC), a plug-in module that calibrates auxiliary-behavior representations before fusion in multi-behavior recommendation. For each auxiliary behavior, TNC estimates a reference subspace from the leading left singular vectors of the auxiliary–target cross-covariance, averaged across paired training users and items. By orthogonally decomposing auxiliary representations into this reference subspace and its null-space complement, TNC learns bounded, entity-specific scalar gains to separately rescale both components under target-ranking supervision. The module is jointly trained with behavior encoders, with the reference projectors recomputed at every optimization step and detached from backpropagation. Theoretical results characterize TNC as a minimal constrained adjustment with optimal cross-covariance energy retention and conditional stability against reference perturbations. Experiments evaluate six backbones on three purchase-recommendation datasets, with additional component and group-wise analyses.

Critical strengths

S1. Clear motivation and well-formulated decoupled calibration. The paper presents a clear and intuitive formulation of auxiliary representation calibration through orthogonal subspace decomposition and null-space constraints. Decoupling the scaling of target-correlated evidence from orthogonal residual evidence directly addresses the limitation of uniform rescaling or hard subspace truncation in prior projection-based MBR methods.

S2. Informative ablations supporting the calibration design. Tables 3 and 4 provide useful empirical evidence for the roles of the target reference and learnable component gains. On Taobao, replacing the cross-covariance reference with auxiliary PCA (w/o Target Reference) reduces NDCG@10 from 0.1070 to 0.0956, close to the uncalibrated baseline of 0.0955. Moreover, removing the null-space component or fixing its gain at one substantially degrades performance.

S3. Precise and well-scoped theoretical analysis. The minimum-change geometric characterization is sound for fixed inputs, references, and gains, and the authors carefully delineate its boundary conditions—explicitly noting that the calibration null space differs from that of the full cross-covariance and that step-wise projection preservation does not imply score invariance or convergence under joint training.

Critical weaknesses

W1. The main comparison is confounded by the prediction interface. Section 5.1 and Appendix B.2 state that +TNC replaces native aggregation or scoring with mean fusion and inner-product scoring where necessary, while Base retains the original predictor. Consequently, for backbones with specialized fusion architectures—most notably SpectraMB and UIPL—+TNC shows negligible gains or slight degradation (Table 2). Conversely, for CRGCN and MBGCN , it is impossible to disentangle how much of the improvement stems from TNC itself versus replacing their native scoring/aggregation interface with mean fusion. A controlled comparison where Base also uses the exact same mean-fusion interface (or where TNC is plugged directly into the native fusion layer of SpectraMB/MBGCN) is missing for the five non-RMBRec backbones.

W2. Insufficient validation of the shared user–item feference. Equation (4) equally averages user- and item-side cross-covariances to construct a shared projector. Appendix A.1 acknowledges that the two association structures may reinforce or cancel in this average, but no ablation compares the shared reference with separately estimated user and item projectors. Such a comparison would help justify this design choice and assess whether a shared reference adequately captures the potentially different association structures on the two sides.

W3. Training cost is not measured. The reference is recomputed at every optimization step, requiring an additional encoder forward pass over all paired entities with dropout and augmentation disabled, followed by cross-covariance estimation and SVD. For graph backbones this adds up to one extra full-graph pass per step, and the paper notes the pass is avoidable only when forward settings coincide. The paper provides no wall-clock training time comparison or investigation into periodic/lazy reference updates.

Questions for authors

Q1. Isolating calibration gains. For backbones whose native prediction interface is replaced, could you compare models with and without TNC under the same mean-fusion and inner-product interface, keeping the remaining training settings identical? This would isolate calibration’s contribution from that of the interface change.

Q2. Training overhead and reference updates. Could you report wall-clock training time per epoch for Base and +TNC under the same hardware and batch settings? On a representative configuration, how does updating the reference projectors every (k) steps, rather than every step, affect training time and recommendation accuracy?

Q3. Rank selection and sensitivity. Could you report the selected rank r∈{4,8,16,32,48}r \in \{4,8,16,32,48\} for each dataset–backbone configuration? For representative configurations, could you also show how NDCG@10 varies with rr and report the fraction of cross-covariance energy retained for each auxiliary behavior, ∑j=1rσj(Cb)2∑j=1dσj(Cb)2\frac{\sum_{j=1}^{r}\sigma_j(C_b)^2}{\sum_{j=1}^{d}\sigma_j(C_b)^2}? These results would clarify the robustness of TNC to rank selection and how the selected ranks relate to the observed association spectra.

Reviewer_d6sWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a plug-in module called target-referenced null-space calibration (TNC). TNC splits each auxiliary behavior representation into a target-associated component and its orthogonal complement. The experimental results show the effectiveness of the proposed TNC module.

Critical strengths

S1. As a plug-in auxiliary representation calibration module, TNC can use potentially complementary preference evidence for target prediction.

S2. The experiments are conducted on three real-world datasets.

Critical weaknesses

W1. It seems that the novelty of the paper is limited. For example, MCLMR wraps existing multi-behavior backbones with a causal intervention module [1]. It is better to review and discuss similar plug-in modules.

W2. This paper lacks the theoretical guidance for hyperparameter selection. For example, the paper provides no theoretical justification for the selection of the dimension r of the target-associated subspace, i.e., r needs to be manually tuned on the validation set. Then it fails to guide hyperparameter selection under different scenarios.

W3. SOTA baselines are missing. There are some SOTA baselines, such as [1][2], which should be carefully compared. It is difficult to evaluate the novelty of the proposed methods without a thorough comparison with the SOTA baselines.

W4. The experiments on time costs should be reported. The null-space constraint restricts nothing at training time. The authors are encouraged to report the related experimental results.

[1] MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation. WWW 2026

[2] Target-Enhanced Gated Transformer: A Multi-Behavior Recommendation Framework for Noise Suppression and Target Signal Preservation. ICMR 2026

Questions for authors

W1-W4

1 comment

Thank the authors for their rebuttal. After carefully reading both the rebuttal and the revised paper, I tend to keep my scores.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.