acceptodds
Under review as a conference paper at ICLR 2027

Training Encoders Need Not Be the Best Readouts: Personalized Post-Training Encoder Adaptation for Collaborative Filtering

12 views

Abstract

Collaborative filtering (CF) learns user and item representations from interaction histories and ranks items by dot-product scores. From matrix factorization to graph-based models, a fixed encoder matrix maps a trainable embedding table to these representations, and the same encoder is reused at inference. Yet the encoder plays a dual role: during training it both constructs representations and shapes gradient updates to the embedding table, whereas at inference it only reads out the converged table. Convergence therefore does not imply that the training encoder is the best readout, and controlled experiments confirm that an alternative encoder can outperform reuse. A user-wise analysis further shows that each encoder row combines collaborative and preference information whose score effects depend on the learned item geometry, calling for a user-specific readout informed by the trained model. Building on this analysis, we propose **PEAR**, *Personalized post-training Encoder Adaptation for Recommendation*, a lightweight framework for linear inference encoders in CF. PEAR freezes the embedding table and final item representations, constructs two encoder bases from training interactions, fitted predictions, and item geometry, and fits at most two bounded coefficients per user, limiting adaptation capacity without learning new embeddings. Experiments across datasets, backbone encoders, and training objectives show that PEAR consistently improves recommendation performance over converged models while preserving the original dot-product scoring interface and serving cost. Code and data are provided in the supplementary material.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
collaborative filtering, post-training adaptation, parameter-efficient adaptation

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

The paper formalizes the two roles of linear CF encoders during training (representation construction and gradient propagation to embeddings) while showing that they serve only the former role during inference. It demonstrates that reusing the training encoder for inference may not be optimal. Building on this observation, the paper proposes PEAR, a post-training method that refines user representations along two user-specific directions derived from item-side and user-side information, while keeping the learned embeddings and final item matrix fixed.

Critical strengths

[S1] The paper offers an insightful theoretical perspective on CF representation learning. Its proposition and counterexample clarify why reusing the training encoder at inference may be suboptimal, making the contribution well aligned with ICLR.

[S2] BPR, DirectAU, and SCCF are representative approaches in collaborative filtering. The paper achieves noteworthy performance improvements by applying the proposed method to these approaches. Furthermore, the paper reports results over multiple independent runs, which makes the experimental results reliable.

Critical weaknesses

[W1] The main text lacks self-containment and relies excessively on the appendix. Furthermore, the overall presentation lacks clarity, making it difficult to fully understand the paper's key ideas from the main text alone.

[W2] The theoretical analysis relies on the assumption that the encoder H\boldsymbol{H} is fixed and linear throughout training. Although many representative collaborative filtering methods satisfy this assumption, it may limit the generalizability of the proposed approach to more diverse encoder architectures.

[W3] Although framed as "encoder adaptation", the proposed method merely adjusts user representations along two predefined directions, making the encoder adaptation narrative somewhat overstated.

[W4] Three of the four evaluation datasets are book-related, while the gains on Beauty are relatively modest. This raises the concern that the method may work well primarily in the book domain.

Questions for authors

[Q1] I believe the main text requires substantial revision for the camera-ready version. Too many essential details (including the definitions of key symbols) are deferred to the appendix, while the explanations in the main text are insufficient to provide an intuitive understanding of the proposed method. Furthermore, some symbols are presented in ways that suggest overlapping meanings despite serving distinct purposes. For instance, Cu\boldsymbol{C}_u and Cm\boldsymbol{C}_m play fundamentally different roles, yet their notation and presentation may cause confusion. What are the authors' thoughts on these concerns?

[Q2] The proposed method appears to modify only the user-side encoder, whereas Figure 1 seems to modify the entire encoder, including the item side. If so, could the authors explain this discrepancy and the rationale behind PEAR's design? Can Figure 1 still be considered a valid motivation for the proposed method?

[Q3] Could the authors also report Recall@k? Since Recall@k can be more important than ranking metrics in practical recommendation systems, this comparison would be valuable.

Reviewer_Pq1QOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits the convention of sharing the encoder between training and inference in collaborative filtering, proposing a lightweight post-training user-side encoder adaptation scheme that decomposes inference performance into collaborative and preference information along two basis directions with personalized scalar coefficients. While the approach demonstrates significant and consistent improvements across various backbones and objectives, the paper would benefit from more analytical experiments, particularly regarding the positioning of the proposed method relative to the baseline schemes in Figure 1, the distribution of personalized coefficients across user activity groups, and broader ablation settings beyond the MF backbone configuration.

Critical strengths

1, This paper revisits the convention of sharing the encoder between training and inference in collaborative filtering models. Through empirical experiments and theoretical analysis, it points out that this encoder sharing is suboptimal, and explores a post-training user-side encoder adaptation scheme for inference.

2, The paper decomposes the user-side recommendation performance at inference into collaborative information from other users and preference information from items. It adjusts the user-side encoder weights along two basis directions corresponding to users and items, and learns two scalar coefficients per user to achieve personalized adaptation.

3, Experimental results show that the proposed method achieves significant performance improvements across various backbone and training objective combinations. Further analyses verify its effectiveness and complementarity compared to existing structural postprocessors approaches, the contribution of each component, and the efficiency and lightweight nature of the overall framework.

Critical weaknesses

1, The empirical study in Figure 1 shows the performance comparison of using different inference encoders versus reusing the training encoder, illustrating the research motivation. However, this motivation is not further discussed in the experiment section, such as where the proposed method's performance would be located in Figure 1.

2, The analysis of personalized adaptation in the experiment section is insufficient. The ablation study only analyzes the performance difference between using personalization or not. A further analysis on the distribution of personalized scalar coefficients, i.e., cuI,cuUc_u^I, c_u^U, across user groups with different interaction activities would better support the benefits of personalization.

3, The ablation experiments are mainly based on the Douban-book with MF setting. More experimental settings are necessary.

4, The overall structure of the paper needs optimization. Preliminaries and Analysis take up too much space, leaving only about 4.5 pages for the method and experiments, which causes some important analytical experiments to be placed outside the main text.

5, Typo: Page 4, Line 176, "app:dual-role".

Questions for authors

1, Where would the proposed method's performance be located in Figure 1?

2, What is the distribution of personalized scalar coefficients cuI,cuUc_u^I, c_u^U across user groups with different activity levels?

3, If using other backbone settings, such as LightGCN, would the ablation trends be similar?

Reviewer_UhiQOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper revisits encoder reuse in linear collaborative filtering. PEAR, a post-training framework that freezes the learned embedding table and final item representations and adapts only the user-side inference readout. PEAR constructs two fixed encoder bases—one recombining user-origin collaborative information and one recombining item-origin preference information. Experiments cover MF and LightGCN backbones, BPR/DirectAU/SCCF objectives, Amazon and Douban datasets.

Critical strengths

The paper evaluates multiple backbones, objectives, datasets, structural postprocessors, generic adaptation controls, and resource usage.

Critical weaknesses

1.The contribution is incremental relative to existing post-training and encoder-substitution work. The paper’s central observation—that the training encoder need not be the best inference readout—is already present in the structural postprocessing literature, e.g., TAG-CF, LightGODE, and related work on decoupling training-time and inference-time propagation. PEAR’s technical contribution is closer to a carefully regularized per-user fine-tuning scheme than to a new learning problem or a new class of encoders. The paper does not establish a conceptual separation from existing post-hoc adaptation methods beyond the specific two-direction parameterization.

  1. The method is limited to linear encoders, and the extension to nonlinear CF is not addressed. Many modern CF backbones use nonlinear transformations, attention, or state-dependent propagation. The main method and theory do not cover nonlinear encoders. Since the paper claims a general post-training framework, the scope should be stated more clearly. The current framing may overstate generality.

  2. No comparison with test-time adaptation and related post-hoc methods. The related work discusses TTEN, EDGE, DT3OR, and TTA-GREC, but the experiments do not compare against them. Without such comparisons, the paper cannot claim that PEAR is a strong or preferable post-training adaptation framework.

Questions for authors

1.Novelty relative to structural postprocessing. The observation that encoder reuse is not always optimal is already central to TAG-CF and LightGODE. What can PEAR express that a structural postprocessor followed by a simple per-user calibration cannot?

2.Test-time adaptation baselines. Why are TTEN, EDGE, DT3OR, and TTA-GREC not compared?

3.Nonlinear encoders. PEAR is formulated for linear encoders. How would it extend to nonlinear or attention-based CF? Is the method fundamentally limited to linear readouts?

2 comments

Dear Reviewer UhiQ,

Thank you again for your review. We would be grateful if you could revisit our responses and revised manuscript, including the additional comparisons and clarified scope of PEAR. Please let us know if any concerns remain—we would be happy to provide further explanations and discuss them with you.

Best regards,
The authors

AuthorsOpenReview ↗

Follow-up on Our Rebuttal Responses

Dear Reviewer UhiQ,

With only one day remaining in the rebuttal period, we would like to kindly bring our responses to your attention. We have provided explanations for all the concerns you raised and clarified the corresponding revisions in our response and revised manuscript. We would be very grateful if you could take a moment to review them.

If any concerns remain or if you have any further questions, please let us know. We would be happy to provide additional explanations and discuss them with you.

Thank you for your time and consideration.

Best regards,
The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.