acceptodds
Under review as a conference paper at ICLR 2027

Warmth Propagation: A Lightweight Content-Aware Approach to Cold-Start Recommendation

Abstract

In this paper we propose Warmth Propagation, a novel approach to mitigating the cold-start problem in recommender systems. It relies on content data to compute pairwise item similarity, and then propagates collaborative information from warm items to similar cold ones, i.e. to items that have no click data. The important feature of the proposed approach is its efficiency: it computes pairwise similarity in an implicit way and hence can deal with large-scale item catalogs. Warmth Propagation is not a stand-alone approach and can be applied to pretrained recommenders that were not originally designed to handle cold items, requiring only that item embeddings be initialized from content features. We experimentally show that the proposed approach introduces minimal computational overhead and outperforms various cold-start recommenders.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
recommender systems, cold start, linear models, transductive learning

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

This paper proposes Warmth Propagation, a post-hoc method for item cold-start recommendation that transfers collaborative information from warm items to cold items through content-based relationships. The method first learns an item reconstruction matrix from content embeddings using an EASE-style objective, then transfers these reconstruction relationships to collaborative embeddings through a regularized least-squares formulation. By exploiting the low-rank structure of the content matrix and the Woodbury identity, the authors derive an exact closed-form solution that avoids constructing item-by-item matrices and scales linearly with catalog size for fixed embedding dimensions. Experiments on three datasets with SASRec and BERT4Rec report improvements in cold-item recommendation while largely preserving overall performance.

Critical strengths

S1. The post-hoc formulation addresses a practically relevant constraint. The method can improve cold-item representations without modifying the recommendation backbone or retraining it with a specialized cold-start objective, provided that suitable content-initialized embeddings are available. This makes the approach potentially useful for adapting existing recommendation systems.

S2. The scalable closed-form solution is a useful technical contribution. The authors exploit the structure of the content reconstruction problem to avoid explicit item-by-item matrices, reducing the main computational dependence to matrices whose dimensions are determined by the content embedding size. This offers an attractive alternative to directly solving large-scale embedding propagation problems.

Critical weaknesses

W1. The methodological novelty over existing embedding reconstruction and propagation techniques is limited. The two optimization stages build directly on EASE, LLE, graph regularization, and ImplicitSLIM, all of which are acknowledged. The main additional contribution appears to be their combination with content embeddings and an efficient algebraic solution. Furthermore, closely related work on content-based collaborative information transfer is missing from the discussion, particularly Content-based Graph Reconstruction for Cold-start Item Recommendation (Kim et al., SIGIR 2024) and Proxy-based Item Representation for Attribute and Context-aware Recommendation (Seol et al., WSDM 2024). The authors should clarify the distinctions from these methods and more precisely identify the methodological contribution beyond existing reconstruction formulations.

Kim, J., Kim, E., Yeo, K., Jeon, Y., Kim, C., Lee, S., and Lee, J. Content-based Graph Reconstruction for Cold-start Item Recommendation. SIGIR 2024.

Seol, J., Gang, M., Lee, S.-G., and Park, J. Proxy-based Item Representation for Attribute and Context-aware Recommendation. WSDM 2024.

W2. The experiments do not sufficiently establish the claimed scalability advantage or the significance of the performance gains. On Beauty, exact kernel regression achieves 0.057 cold NDCG@10 compared with 0.058 for Warmth Propagation, suggesting that computational efficiency may be the primary advantage over this strong baseline. However, no systematic wall-clock time, peak memory, or catalog-size scaling results are reported to substantiate the claimed negligible overhead. The results also lack variability across training seeds or data splits. Moreover, the method uses a cold-item-focused validation objective, and it is unclear whether the competing methods receive comparable hyperparameter optimization. These issues make the practical advantage difficult to assess conclusively.

Questions for authors

See weaknesses.

Reviewer_R4AhOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Warmth Propagation (WP), a post-hoc method for addressing item cold-start in recommender systems. WP aggregates collaborative signals from multiple content-similar warm items to reduce the impact of noise in individual embeddings. The authors evaluate the method on several benchmarks using NDCG@10 and conduct separate evaluations for different groups of cold items.

Critical strengths
  1. The proposed method is easy to deploy as a post-hoc plug-in, requiring no additional training or retraining of the underlying recommender. Its computational complexity scales linearly with the number of items, making it practical for large item catalogs.

  2. The method constrains updated warm-item embeddings to remain close to their original learned representations, aiming to preserve warm-item recommendation performance while improving performance on cold items.

  3. The experiments compare the method across multiple datasets and include dedicated evaluations on cold items, providing a clearer assessment of its effectiveness in the cold-start setting.

Critical weaknesses
  1. Limited novelty. ImplicitSLIM already explores a closely related formulation involving an EASE/SLIM-style objective for learning an item–item weight matrix and an LLE-style objective. The conceptual contribution therefore appears incremental, and the paper should more clearly articulate its methodological advances over this prior work.

  2. Limited evaluation metrics. The evaluation relies exclusively on NDCG@10. Reporting additional metrics, such as Recall, and results at multiple cutoff values would provide a more comprehensive assessment of recommendation quality and help establish whether the improvements are consistent.

  3. Limited baseline coverage. The evaluated baselines are relatively simple and mostly older methods. Comparisons with more recent, relevant approaches to item cold-start recommendation would strengthen the empirical evaluation and better establish the method’s competitiveness.

Questions for authors
  1. How would WP perform against a ridge regression baseline that maps content features to collaborative embeddings for both warm and cold items, while blending the mapped warm-item embeddings with their original embeddings EE?

  2. Do the linear regression, MLP, and approximate kernel regression baselines in Table 2 update warm-item embeddings, or only cold-item embeddings?

  3. In Table 3, why does cold-only WP (0.015) perform worse than plain content initialization (0.018)? This raises concerns about the benefit of propagation alone. Could the authors clarify how much of the overall improvement comes from propagation versus updating warm-item embeddings?

  4. The paper highlights efficiency as a key advantage, but I could not find experiments supporting this claim. Could the authors report runtime and memory usage, compare these with relevant baselines, and evaluate how the computational cost scales with the number of items?

1 comment

We thank the reviewer for their careful reading of our paper, thoughtful comments, and constructive feedback. We address the raised points below.

W1: The objectives of the proposed approach are indeed similar to those of ImplicitSLIM. However, in the considered setting, we employ content similarity rather than collaborative similarity, which allows us to make the Step 2 objective of ImplicitSLIM simpler and more natural. Moreover, the analytical solution of WP is completely novel (except for the part where both ImplicitSLIM and WP build upon the analytical solution of EASE) and involves no approximations, unlike ImplicitSLIM. We consider the proposed analytical solution to be one of the most important contributions of the paper.

Q1: We performed this experiment with all baseline mappers (linear regression, MLP, and approximate kernel regression) but did not report the results because they were very close to those obtained when updating only cold-item embeddings. Thus, unfortunately, interpolation between predicted and original warm-item embeddings does not resolve the problem discussed in Section 5.3.

Q2: Yes, both warm-item and cold-item embeddings are updated here. As we reported in Section 4.3, updating only cold-item embeddings leads to near-zero performance on cold items.

Reviewer_J5vKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Warmth Propagation, a post-hoc method for item cold-start recommendation. It learns a content-space reconstruction on the content embedding matrix, then transfers this geometry to the collaborative embedding space. Experiments use SASRec and BERT4Rec backbones on three datasets, and compare against baselines from (Pembek et al., 2025).

Critical strengths
  1. Casting cold-start embedding construction as geometry transfer from content space to collaborative space is nice.
Critical weaknesses
  1. The baselines are from only one paper (Pembek et al., 2025) in Table 2.
Questions for authors

increasing more baselines from more papers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.