acceptodds
Under review as a conference paper at ICLR 2027

Multi-Positive Top- Recommendation: A Differentiable Dynamic Programming Approach

Abstract

Modern recommender systems typically optimize for top- retrieval; however, they largely rely on continuous surrogate loss functions due to the non-differentiable nature of ranking metrics. A fundamental limitation of these conventional losses is that they evaluate positive items independently—contrasting a single positive item against one or more negative items per instance. Consequently, they fail to capture the joint combinatorial structure and mutual associations among a user's multiple positive interactions. To overcome this bottleneck, we adopt a novel framework powered by Differentiable Dynamic Programming (DiffDP), which reformulates multi-positive item selection as a dynamic programming process. Specifically, we leverage the Fenchel-Young loss framework—a theoretically rigorous surrogate objective that enables direct optimization over joint multi-positive subsets. By regularizing the multi-positive selection problem with Shannon entropy, our dynamic programming formulation recovers the exact log-partition function of a Conditional Poisson Binomial Distribution (CPBD). Unlike Gumbel-perturbed relaxes or sampling-based approaches, our formulation models simultaneous unordered sets under a strict global capacity constraint. This key property enables the exact, deterministic, and noise-free computation of marginal inclusion probabilities via automatic differentiation. Extensive experiments on real-world benchmark datasets demonstrate that our approach yields competitive top- recommendation performance. The code is provided anonymously at https://anonymous.4open.science/r/Differentiable-Dynamic-Programming-1857.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-Positive, Top-$K$ Recommendation, Differentiable Dynamic Programming

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

The paper proposes DiffDP, a multi-positive recommendation objective based on entropy-regularized dynamic programming and the Fenchel–Young loss framework. It computes the log-partition function and inclusion marginals of a fixed-cardinality subset distribution in (O(MP)). Experiments across four datasets and three recommendation backbones show generally improved ranking performance.

Critical strengths
  1. The DP formulation provides exact marginal gradients for the specified subset distribution and integrates naturally with existing recommendation models.

  2. The evaluation covers four datasets, three backbones, and a diverse set of baseline losses. The broadly consistent improvements support the practical potential of the objective.

Critical weaknesses
  1. The claim about capturing "joint item associations" seems too broad. In Sec. 3.3, the subset score is still a sum of individual item scores, and fixing PP introduces competition between items. I do not see how this also captures complementary relationships among positive items.

  2. I found the definition of ΩP\Omega_P in Sec. 3.2 unclear. Entropy of the marginals and joint entropy over subsets are different quantities. The paper should specify the regularizer for which the DP log-partition is the convex conjugate. Calling it negative Shannon entropy over the marginal polytope leaves this ambiguous.

Questions for authors
  1. What does joint associations cover beyond the dependence introduced by fixing the subset size?

  2. Could the authors give the explicit definition of ΩP\Omega_P used in Eq. (7) and explain its connection to the DP log-partition?

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

This paper proposes a new Top-KK recommendation loss, DiffDP. In contrast to traditional Top-KK recommendation losses, which only consider the single positve item with negative items, DiffDP focus on multiple postive items with negative items.

Critical strengths
  1. The Idea of this paper is fascinating (although the same idea in [1]), especially on how to use dynamic programming to construct the algorithm.
  2. The selected baselines and experiments are sufficient.

Reference

[1] Vivier-Ardisson G, Sander M E, Parmentier A, et al. Differentiable Knapsack and Top-k Operators via Dynamic Programming[J]. arXiv preprint arXiv:2601.21775, 2026.

Critical weaknesses
  1. The authors claim two challenges for one-hot targets in introduction. However, the authors did not provide sufficient empirical experiments on these two claims: (1) it destroy joint associations and combinatorial structure; (2) it induces “false competition”. The authors are supposed to provide empirical experiments to prove these two points. Please do not just claim DiffDP outperforms than other single-positve item method.
  2. The DP process of the algorithm has to iterate P times. Whether this iteration would significantly increate the training time cost. The authors are supposed to report the training time of one epoch with other single postive item method.
  3. The authors are supposed to conduct Top-KK consistency experiments to verify whether the selective PP truly improve the Top-KK performance.
  4. I read the source code. It seems that the authors use all the negative items, but not sampled negative items to conduct DP. The authors are supposed to clarify this, or I suspect the performance improvement comes from the more negative items.
Questions for authors
  1. The motivation for using a multi-positive item is the lack of empirical experimental support. Please refer to Weakness 1.
  2. The authors are supposed to compare with SL@K [1], which is a SOTA Top-KK recommendation loss.
  3. The authors are supposed to add a section in the Appendix to formulate the state function. What does D[i,j] represent, and how does D[i,j] transfer from D[i-1,j-1] and D[i,j-1]? And the authors are supposed to present the pseudocode of the DP process. I carefully read the source code of this paper, and thus can understand how the algorithm runs. Adding new section in Appendix to demonstrate the DP process would help readers to understand the implementation.
  4. I find that the idea of [2] is very similar to this paper. The authors are supposed to clarify this to demonstrate the novelty of this paper. Although the authors have discussed this paper in the related work, the authors did not compare DiffDP with this paper.

The idea of this paper (especially the DP process) is very fascinating for me. However, the motivation for why the authors use multiple positive items is not clear, and the implementation setting is not clear (see Weakness 4). Furthermore, I find that the idea of [2] is very similar to this paper. What is the difference between [2] and this paper? If the authors could address these questions aforementioned, I strongly would like to raise my score.

Reference

[1] Weiqin Yang, Jiawei Chen, Shengjia Zhang, Peng Wu, Yuegang Sun, Yan Feng, Chun Chen, and Can Wang. 2025. Breaking the Top-K Barrier: Advancing Top-K Ranking Metrics Optimization in Recommender Systems. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD '25).

[2] Vivier-Ardisson G, Sander M E, Parmentier A, et al. Differentiable Knapsack and Top-k Operators via Dynamic Programming[J]. arXiv preprint arXiv:2601.21775, 2026.

Reviewer_UnW3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores how recommendation models can learn from several positive items in a user’s history at once. The authors aim to bring the training objective closer to the practical goal of recommending multiple items together. Computing these probabilities directly would require considering an enormous number of combinations. The authors address this with differentiable dynamic programming, which tracks whether each item is included and how many items have been selected. This allows the normalization term to be computed exactly without enumerating every possible set. Entropy smoothing and the Fenchel–Young framework then yield a training loss whose gradient for each item’s score is its predicted inclusion probability minus its binary target label.

The authors evaluate the method on standard benchmarking datasets. They report improvements over losses such as BPR, Softmax, and Talos in most settings. Using more positive items per training example generally improves recommendation performance, although it also increases training times.

Critical strengths
  1. The method has a clear probabilistic interpretation and can be used with existing recommendation models. The dynamic program computes the normalization term and item inclusion probabilities exactly, without Monte Carlo estimates. It changes the training loss while keeping the model’s scoring and inference procedures unchanged, making it relatively straightforward to integrate.
  2. The experiments cover a useful range of settings. The authors evaluate four datasets and three recommendation backbones against several losses, including Talos, and report averages over five random seeds. The improvements in most settings suggest that the objective is worth studying further.
Critical weaknesses
  1. I am not yet clear on what the paper adds beyond existing work. Differentiable Top-k dynamic programming, entropy smoothing, and the corresponding Fenchel–Young losses have already been developed. The authors need to explain what is new beyond their application to recommendation. The motivation also overstates the limitations of conventional losses: BCE supports multi-hot targets, so it should not be described as inherently restricted to one-hot labels.

  2. The experiments do not separate the benefit of subset modeling from the benefit of using more positives per update. Increasing P also provides more positive supervision. A useful control would apply ordinary losses to the same P positives and sum or average them before comparing with DiffDP. There is also a discrepancy between the paper and the released code: the paper specifies 128 sampled negatives and positive sampling with replacement, while the code uses full-catalog scores and samples positives without replacement. These choices change the training conditions, so the authors should clarify which protocol produced the main results.

Questions for authors
  1. What does the method add to existing differentiable Top-k DP and Fenchel–Young losses, either algorithmically or theoretically? Also, since BCE supports multi-hot targets, what exactly do you mean by the “single-positive limitation”?
  2. Did all methods see the same number of positive examples and receive the same number of training updates? Do you have a comparison where ordinary losses use the same P positives, with their losses summed or averaged, and how does DiffDP perform against that control?
  3. The paper specifies 128 sampled negatives and positive sampling with replacement, but the released code uses the full catalog and samples positives without replacement. Which setup produced the main results? How were users with fewer than P positive items handled?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.