acceptodds
Under review as a conference paper at ICLR 2027

Robust Recommendation under Unknown User Intent via Worst-Objective Optimization

Abstract

Modern recommender systems must simultaneously serve multiple user objectives, such as relevance, engagement, diversity, novelty, and serendipity. A common solution is to optimize a fixed objective or a user-specific scalarization of several objectives. However, even a personalized scalarization is typically fixed for a user, while the user's immediate intent can change across visits and even within a session. A policy that is excellent for the dominant intent can therefore fail severely when a less frequent intent becomes active. We study recommendation under an unknown, time-varying intent distribution and propose Objective-Diverse Recommendation (ODR), a robust learning principle that directly optimizes the weakest objective for each user rather than committing to a fixed scalarization. Our formulation views the unknown intent distribution as an adversary over a set of candidate objectives and reduces to a max–min ranking objective. We provide a policy-gradient procedure that, at each update and for each context, selects the currently worst-performing objective as the reward, thereby estimating the gradient of the max–min objective without requiring the true intent distribution at deployment. Synthetic experiments with relevance, diversity, novelty, and serendipity objectives, together with experiments on the KuaiRec full-feedback benchmark, show that the proposed approach maintains stable worst-objective performance across substantial shifts in the intent mixture. Its advantage becomes more substantial as the number of plausible intents increases. These results suggest that robust optimization over plausible objectives can be effective even when the user’s active intent distribution is unknown and not explicitly estimated.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
reward design, metric design, recommendations, long-term objective

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dmVLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ODR, which optimizes the worst-performing objective for each context, measured as relative improvement over the logging policy, and trains by using that objective as the policy-gradient reward. Experiments on a 10-item synthetic environment and KuaiRec report the best worst-objective improvement across intent mixtures.

Critical strengths

The objective is simple, clearly described, and requires no change at serving time.

Critical weaknesses

W1. The KuaiRec experiment seems to favor ODR by design. For each user, ODR selects the candidate ranking that maximizes the worst-objective score and is then evaluated on the same metric. Its advantage over the baselines is therefore largely expected and does not necessarily demonstrate its effectiveness in real-world recommendation.

W2. The core idea is to identify the worst-performing objective during training and update the model accordingly. However, the KuaiRec experiment simply selects from pre-generated rankings, while the synthetic experiment directly optimizes a differentiable surrogate rather than using the proposed policy-gradient procedure. Thus, the proposed training algorithm is not directly evaluated, leaving its practical effectiveness unclear.

W3. Relevance is typically the primary objective in recommendation systems. It remains unclear how much relevance ODR sacrifices to improve weaker objectives and whether this trade-off is worthwhile.

Questions for authors

Q1. Is ODR's superior performance due to effective learning, or simply because it directly selects rankings that maximize the evaluation metric?

Q2. What happens if a recommendation model is actually trained using the proposed ODR algorithm and evaluated on unseen users and items, rather than selecting from 45 pre-generated rankings?

4 comments

We sincerely thank the reviewer for the constructive feedback and for recognizing the practical motivation and potential value of our approach. We appreciate the opportunity to clarify our experimental design and the intended scope of our empirical evaluation.

1. Does optimizing the evaluation metric favor ODR?

We thank the reviewer for raising this question. However, we respectfully disagree that alignment between the optimization objective and the evaluation criterion constitutes a methodological weakness.

Our primary research question is whether explicitly protecting the weakest plausible objective provides a useful alternative to optimizing a single objective or a fixed scalarization. Accordingly, evaluating worst-objective performance is the direct and appropriate way to assess whether this robustness principle achieves its intended goal.

The same reasoning applies to conventional recommendation methods. Relevance-oriented methods are evaluated using relevance metrics, while diversity-oriented methods are evaluated using diversity metrics. Evaluating an algorithm against the objective it is designed to optimize is not inherently problematic, provided that the resulting claims are interpreted appropriately.

Importantly, we do not rely exclusively on worst-objective performance. Figures 2(b) and 4(b) additionally report individual metric improvements, allowing readers to examine the trade-offs between relevance and discovery-oriented objectives.

We therefore believe that the experiment provides a valid comparison of alternative optimization objectives within the specified candidate class. It is not intended to establish causal superiority in real-world user satisfaction, and such a claim is not necessary to support our central contribution.

2. Training an actual recommendation model

We thank the reviewer for this question. We believe the concern stems from a misunderstanding of the role of candidate generation and re-ranking in our experiments. Our method is not intended to learn the candidate generation process itself. Rather, ODR optimizes a parameterized re-ranking policy after candidate generation, which is a standard and practically important stage in recommendation systems.

In the KuaiRec experiments, we construct a common candidate-ranking pool and optimize a neural re-ranking policy according to the proposed worst-objective. Thus, the candidate generation process is fixed, but the final re-ranking decisions are optimized using our proposed objective, rather than simply adopting a predetermined ranking. This design intentionally isolates the effectiveness of the proposed optimization principle from differences in candidate generation.

We therefore respectfully disagree that our experiments merely evaluate pre-generated rankings without optimizing a recommendation policy. Nevertheless, we recognize that the distinction between candidate generation and re-ranking may not have been sufficiently clear in the current manuscript, and we will clarify this experimental setup to avoid misunderstanding.

3. Relevance–robustness trade-off

We thank the reviewer for emphasizing the importance of relevance. However, we would like to point out that the trade-offs between relevance and other objectives are already explicitly examined in our experiments.

Figures 2(b) and 4(b) report individual metric improvements, including relevance-oriented metrics. These results allow readers to assess how ODR balances relevance against diversity, novelty, and serendipity, rather than evaluating the worst-objective metric in isolation.

Moreover, the possibility of sacrificing some relevance to protect weaker objectives is not an unintended limitation, but an inherent and deliberate property of the robust optimization formulation. Our objective is designed for settings in which severe failures on less frequent user priorities are considered costly.

The relative importance of relevance and robustness ultimately depends on the application's utility structure. Importantly, our proposed objective can be easily extended to accommodate different application-specific preferences, for example, by imposing a minimum relevance constraint or optimizing a weighted combination of relevance and worst-objective performance. Such extensions allow practitioners to flexibly control the trade-off between relevance and robustness. However, no specific threshold, weighting scheme, or alternative utility criterion is suggested in the review.

Without an explicit criterion for determining an acceptable trade-off, it is difficult to conclude that the observed relevance sacrifice constitutes a methodological weakness. Our experiments already expose the relevant trade-offs, while the proposed formulation provides a principled alternative for applications prioritizing protection against severe objective-level failures.

We would like to kindly remind the reviewer that we have submitted a revised manuscript, with all changes highlighted in red, addressing the key concerns raised in the review. In particular, we have clarified the neural re-ranking policy learning setup, the rationale for evaluating worst-objective robustness, and the relevance–robustness trade-offs already examined in our experiments.

We would appreciate any remaining questions or substantive objections before the discussion deadline, so that we can address them directly. If no further concerns are raised, we believe the major issues identified in the review have been satisfactorily resolved and hope that these revisions will be reflected in a more positive final assessment.

Thank you for your response. Since my main concerns remain unaddressed, I am keeping my score.

W1. I do not object to evaluating a method on its target criterion. However, on KuaiRec, ODR selects the ranking that maximizes the worst-objective score directly on the same users it is evaluated on. This advantage is therefore guaranteed by construction and says little about whether the underlying learning principle actually generalizes.

W2. The response states that a neural re-ranking policy is trained, whereas the paper describes the KuaiRec experiment as selecting the best pre-generated ranking per user—please clarify which is accurate. If a policy is indeed trained, performance on unseen users (Q2) would be the key evidence needed. Additionally, the synthetic experiment aspect of W2 remains unaddressed: because it optimizes a differentiable surrogate, the proposed policy-gradient procedure is still untested.

W3. The figures show that a trade-off exists, not that it is worthwhile, and the drop in relevance appears substantial. The λ\lambda-weighted variant mentioned in the response has not been evaluated; plotting a relevance vs. worst-objective trade-off curve for this variant would directly address this concern.

We sincerely thank the reviewer for the follow-up comments. However, we believe that the remaining concerns largely stem from misunderstandings of our experimental protocol and the central research objective of ODR. We clarify these points below.

W1. Generalization in the KuaiRec experiment

We respectfully disagree with the reviewer's characterization of our experimental setup. In the KuaiRec experiments, users are explicitly divided into training and test sets. The respective recommendation policies are learned using the training users and subsequently evaluated on held-out test users. Therefore, ODR does not optimize rankings using the evaluation feedback of the same users on which its performance is reported. Contrary to the reviewer's assertion, the observed improvement on test users is not guaranteed by construction, but reflects the performance of a learned policy on unseen users. Our experimental protocol thus directly evaluates generalization, and we do not believe the concern raised in W1 applies to our actual experiments.

W2. Learning a parameterized recommendation policy

As we have already clarified in the revised manuscript, ODR learns a parameterized neural re-ranking policy from training data, rather than merely selecting the highest-scoring pre-generated ranking for each evaluation user. Its performance is then measured on held-out users, directly addressing the reviewer's request for evidence of generalization. We therefore respectfully disagree that our experiments fail to evaluate a learned recommendation policy.

Regarding the synthetic experiments, we have explicitly explained that they use differentiable soft-ranking surrogates to implement the same active-worst-objective optimization principle. Although we do not separately benchmark the score-function estimator presented as an alternative gradient implementation, our contribution is the robust optimization formulation and its context-dependent learning principle, not a new score-function estimator. The synthetic experiments already provide empirical evidence for this proposed learning principle, and the absence of a separate score-function benchmark does not invalidate that evidence.

W3. Relevance–robustness trade-off

We believe this concern reflects a misunderstanding of our central research objective. ODR is designed to improve the worst-performing objective under unknown and potentially varying user intent distributions, rather than to optimize the trade-off between relevance and other recommendation metrics. Relevance itself can be included among the plausible objectives, and our formulation protects against poor performance on whichever objective is weakest for a given user.

We previously discussed a relevance-weighted extension in response to the reviewer's interest in controlling this trade-off. However, this extension is not the motivation or primary contribution of our paper, and evaluating its trade-off curve would address a different research question. Our existing experiments already report individual metric performance to make the consequences of worst-objective optimization transparent. We therefore do not believe that the absence of experiments on an optional relevance-weighted variant constitutes a limitation of the proposed method.

Overall, we believe that W1 and the principal concern in W2 are addressed by clarifying our actual learning and held-out evaluation protocol, while W3 requests an analysis beyond the explicitly stated objective of our work. We respectfully ask the reviewer to reconsider these concerns in light of the clarified experimental setup and the intended scope of our contribution.

Reviewer_92bXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method: Objective-Diverse Recommendation (ODR), which maximizes the worst improvement across a predefined set of recommendation metrics. The authors treat these metrics as possible user intents whose distributions are unknown, present a policy-gradient formulation, and evaluate the objective using synthetic experiments and a ranking benchmark.

Critical strengths

a) The paper is generally easy to follow and correctly says good population-average metrics can hide poor performance on individual objectives.

b) The max–min formulation is understandable, and the active-objective gradient characterization looks valid.

Critical weaknesses

a) The authors call attributes like relevance, diversity, novelty, and serendipity utilities associated with different user intents. However, these can be interacting properties rather than distinct user goals. How this connects to user intent is difficult to understand from the current manuscript. This leaves a gap between the motivating problem and the chosen objective.

b) The methodological contribution is very limited. The method identifies the weakest objective and follows its gradient using a standard policy-gradient approach. While I can understand the authors' attempt to apply it to a specific case, its additional methodological novelty or practical utility is not at all clear.

c) The experiments do not directly test the score-function estimator. Also, the authors are requested to address the inconsistencies in normalization and scalarization across the baselines.

d) The finite-sample guarantee needs a clearer scope and correction. Authors switch between context-wise minima and across-context objective averages in the theoretical analysis. Specifically, controlling the error in these averages does not imply control over the expected context-wise minimum; hence, the stated guarantee does not follow (Corollary 3). This needs correction or explanation.

Questions for authors

Please refer to the weakness.

6 comments

We thank the reviewer for the careful assessment of our work and for recognizing the relevance of addressing uncertain user intents in recommendation. We appreciate the detailed theoretical feedback and welcome the opportunity to clarify the scope and novelty of our contributions.

1. Relationship between ranking metrics and user intents

We thank the reviewer for raising this conceptual point. Our framework does not require different ranking metrics to correspond to mutually exclusive psychological states or independently identifiable user intents.

Rather, as formalized in Section 3, we assume a finite set of plausible intent-specific utility functions. Relevance, diversity, novelty, and serendipity are used as operational proxies for different possible user priorities, consistent with the multi-objective recommendation literature.

The fact that these metrics can be related or simultaneously desirable does not invalidate the formulation. Our robust objective is defined over a specified set of utility functions, and its mathematical validity does not depend on whether these functions correspond to statistically independent or psychologically distinct intents. Indeed, accommodating multiple potentially competing evaluation criteria without assuming a known preference distribution is precisely the setting that motivates ODR.

While establishing a direct correspondence between individual metrics and latent psychological intents would be an interesting extension, this is not required for the objective-space robustness problem we formulate. In practice, the suitability of a given metric set can be quantitatively evaluated through online A/B testing using downstream outcomes such as user satisfaction and retention, allowing practitioners to refine the metric set based on empirical evidence. We therefore believe that the reviewer's observation concerns the interpretation and practical design of the chosen utility functions rather than a fundamental limitation of the proposed framework.

2. Methodological novelty

We appreciate the comment but respectfully disagree that the methodological contribution is limited to applying a standard policy-gradient procedure to the weakest objective.

The central novelty lies in the formulation of context-dependent recommendation robustness under unknown intent distributions, not in claiming that the mathematical derivative of a pointwise minimum is itself novel.

Specifically, ODR combines three key elements as follows.

  • A per-context max–min recommendation objective motivated by uncertainty over plausible user intents. Reference-policy-normalized improvement, enabling robustness comparisons across heterogeneous ranking metrics.
  • A context-dependent active-objective training rule, where optimization pressure changes across both contexts and learning iterations.
  • An important distinction is that ODR optimizes the expected per-context minimum, rather than the minimum of context-averaged objective values. This difference is fundamental when the weakest objective varies across users or contexts.

Our formulation is also distinct from conventional fixed or personalized scalarization, which commits to a particular weighting of objectives rather than explicitly protecting the weakest plausible objective.

These connections to DRO and multi-objective optimization are already discussed in Section 2 and Appendix A. While individual mathematical components have established precedents, their integration into the specific recommendation problem studied here constitutes the methodological contribution.

If the reviewer is aware of a directly comparable method that optimizes the same context-dependent robustness criterion under unknown intent distributions in the recommendation domain, we would appreciate a specific reference and an explanation of the relevant overlap.

3. Evaluation of the score-function policy-gradient estimator

We thank the reviewer for this observation. We would like to emphasize that the proposed active-worst-objective optimization principle is both theoretically characterized and empirically investigated, although different experimental settings use different implementations.

Section 4.2 presents a score-function policy-gradient estimator for stochastic ranking policies. Under the stated assumptions, this is a valid gradient representation of the selected objective.

In the synthetic experiments, as described in Appendix C.2, we instead use direct differentiation through differentiable soft-ranking metrics. Importantly, this implementation follows the same active-worst-objective gradient principle, allowing us to evaluate the proposed optimization mechanism without introducing unnecessary variance from score-function gradient estimation.

These choices are deliberate. Our primary contribution concerns the robust optimization objective and its context-dependent adaptation, rather than a new variance-reduction technique for score-function estimation.

Although the experiments do not separately benchmark the statistical efficiency of the score-function estimator under logged bandit feedback, we do not claim such a contribution. We therefore believe that this distinction does not invalidate either the theoretical gradient characterization or the empirical evidence supporting the proposed robustness principle.

4. Finite-sample guarantee

We thank the reviewer for carefully examining the theoretical analysis. We agree that the distinction between the expected per-context minimum and the minimum of context-averaged objectives must be handled rigorously.

Importantly, the underlying robustness formulation and the central stability argument remain valid. The minimum operator is 1-Lipschitz with respect to the infinity norm, which provides a direct route to controlling errors in the robust objective.

The relevant bound operates on context-dependent utility estimates. In particular,

\leq \max_z \left| \widehat{G}_z(x;\pi) - G_z(x;\pi) \right|.$$ Consequently, if the expected maximum context-wise estimation error is uniformly bounded by over the considered policy class, the standard empirical-selection argument yields a 2 robust-objective suboptimality bound. For the finite-policy-class, full-feedback setting, a corresponding concentration result can also be derived by applying concentration directly to the bounded per-context robust score, followed by a union bound over the policy class. We acknowledge that Corollary 3, as currently stated, does not sufficiently distinguish context-wise estimation error from error in context-averaged objective values. However, this issue is localized to the assumptions and derivation of the finite-sample bound. It does not affect the minimax equivalence, the active-objective gradient characterization, or the empirical results. We therefore do not believe that this correctable technical issue weakens the central methodological contribution of ODR. ## 5. Normalization and consistency of scalarization comparisons We thank the reviewer for this point. We would like to emphasize that the normalization procedure and baseline objectives are already explicitly specified in Appendix C.1 and C.3.5. The primary purpose of normalization in ODR is to make heterogeneous objectives comparable in terms of improvement relative to the logging policy. This is an integral component of our robust objective, rather than an incidental evaluation adjustment. The scalarization baselines optimize conventional combinations of raw ranking metrics, reflecting their respective optimization principles. Importantly, all methods select from the same candidate-ranking pool and are evaluated using the same reported performance criteria. Thus, the comparisons are controlled with respect to the available candidates and evaluation protocol. We recognize that applying identical normalization to scalarization baselines would provide an additional ablation separating the effect of normalization from the aggregation objective. However, such a comparison would address a more specific question and would complement rather than replace our current evaluation. We therefore believe that the existing comparisons remain meaningful for evaluating the proposed robust recommendation objective against conventional optimization alternatives. The normalization choices are documented and reproducible, and the review does not identify a specific numerical inconsistency that invalidates the reported findings.

We would like to kindly remind the reviewer that we have submitted a revised manuscript, with all changes highlighted in red, addressing the key conceptual, theoretical, and experimental concerns. Most importantly, we have revised the finite-sample analysis to rigorously distinguish context-wise robust objectives from context-averaged objective values, and have clarified the corresponding assumptions and guarantees. We have also strengthened the explanations of methodological novelty, optimization, and baseline comparisons.

We would particularly welcome any remaining technical objections before the discussion deadline, while there is still an opportunity to respond. Otherwise, we believe that the central concerns raised in the review have been adequately addressed and hope that this substantive theoretical revision will be appropriately reflected in the final assessment.

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

The paper, Robust Recommendation Under Unknown User Intent via Worst-Objective Optimization, proposes a recommendation approach for situations where users may have different, unknown goals, such as seeking relevant content or discovering something new. Instead of optimising average performance across objectives, the method maximises the worst-performing objective, aiming to provide a more balanced recommendation experience. The authors support this formulation with theoretical motivation and experiments on synthetic environments and the KuaiRec dataset.

Critical strengths

Formulating recommendation as a worst-objective optimization problem provides a principled way to address this issue without requiring the intent distribution to be known.

The paper establishes a connection between distributionally robust optimization over unknown intent distributions and a max–min objective over intent-specific recommendation metrics. This gives a clear mathematical justification for prioritizing the weakest objective, rather than introducing a heuristic without a principled rationale. The accompanying analysis also discusses the conditions under which the proposed optimization rule is justified.

The paper evaluates the approach in both controlled synthetic environments and on the KuaiRec benchmark, comparing it with single-objective, weighted-scalarization, and constrained-optimization baselines.

Critical weaknesses

The experimental results demonstrate the behavior of the proposed optimizer in selected settings, but do not establish that it reliably converges to a good solution of the max–min objective. The paper explicitly does not claim global convergence for deep, non-convex policy optimization.

The synthetic environment and action catalog encode trade-offs between objectives such as relevance, novelty, and serendipity. Such a setup is useful for studying the proposed method, but may favor an approach specifically designed to balance conflicting objectives.

The reported comparisons would be more convincing with confidence intervals or other uncertainty estimates for both worst-objective and average performance, obtained across independent random seeds and data splits where applicable.

The main KuaiRec comparison selects from a finite set of candidate rankings using fully observed metrics, rather than learning a policy from incomplete user feedback. Moreover, the experiments do not establish causal improvements in user satisfaction or retention. observational analysis relates dissatisfaction to subsequent churn; however, the recommendation method itself was not shown to cause lower churn.

The main-text normalization (Page 4) divides by a baseline that can be zero; the appendix fixes this with epsilon_0, but the definitions are inconsistent. the paper's central mathematical idea is plausible, but its definitions and implementation claims deserve scrutiny. The normalization issue is a concrete inconsistency, not merely a matter of taste.

Questions for authors

Does the fact that the worst-objective score improves during the reported experiments provide sufficient evidence that the proposed optimization algorithm is effective, or would you require additional experiments? Why?

How could the authors demonstrate that their method generalises beyond the specific synthetic environments they designed? What additional experiments would you recommend?

The paper reports improvements in worst-objective performance, but the strength of the evidence depends partly on how consistent those improvements are across experimental runs and data splits. Can the authors report confidence intervals or variability across multiple random seeds and data splits?

3 comments

We sincerely thank the reviewer for the thoughtful and constructive feedback, and for recognizing the importance of our problem formulation and the potential of worst-objective optimization for robust recommendation. We appreciate the opportunity to further clarify our theoretical and empirical contributions.

1. Effectiveness of the proposed max–min optimization

We thank the reviewer for raising this question. We would like to emphasize that both the theoretical justification and empirical optimization behavior are already investigated in the paper.

First, Theorem 1 in Appendix B.4 characterizes the gradient of the per-context worst-objective criterion. Under the stated assumptions, optimizing the currently worst-performing objective provides the gradient of the proposed robust objective. Thus, the update rule is theoretically justified rather than being an arbitrary heuristic.

Second, Appendix D.2 (Figures 7 and 8) explicitly examines learning dynamics, showing how optimization pressure shifts across objectives and how worst-objective performance evolves during training.

The reviewer raises concerns about convergence in non-convex optimization, but global convergence is neither claimed nor required for our principal contribution. Such guarantees are generally difficult for neural ranking models, and their absence does not invalidate the proposed objective or its gradient characterization.

We therefore believe that the existing theoretical analysis and empirical diagnostics provide meaningful evidence of the optimization mechanism. If the reviewer has a particular convergence criterion or optimization failure mode in mind, we would appreciate a more specific explanation of what additional analysis would be most informative.

2. Generalization beyond the synthetic environment

We appreciate the question. However, the objective conflicts introduced in our synthetic environment are intentional and constitute a fundamental part of the problem being studied, rather than an artificial advantage introduced specifically for ODR. The purpose of this controlled setting is precisely to examine how different optimization principles behave when plausible recommendation objectives conflict.

Importantly, our evaluation is not limited to this synthetic environment. Section 6 additionally reports experiments on KuaiRec using real user–item feedback. We also systematically vary intent mixtures, the number of plausible objectives, and logged-data sizes. Appendix D.1 further investigates sensitivity to ranking length.

These experiments examine ODR from multiple complementary perspectives and demonstrate that the observed behavior is not restricted to a single synthetic configuration.

Although further evaluations would always be valuable, the reviewer does not identify a specific alternative setting or failure mechanism that would challenge our conclusions. We would welcome a concrete suggestion regarding which additional distributional characteristics or experimental conditions are expected to alter the findings. In the absence of such evidence, we believe the existing synthetic and real-data evaluations provide substantial support for our claims.

3. Statistical uncertainty across runs

We thank the reviewer for the suggestion. We would like to clarify that statistical variability is already explicitly reported in the figures throughout our paper, including those in the main text.

As detailed in Appendix C.1, the experimental curves aggregate five independent repetitions, with shaded regions in the figures representing plus/minus one empirical standard deviation. Thus, statistical variability is not merely discussed in the appendix but is directly visualized in the main experimental results. The synthetic evaluation also uses independently sampled test contexts, as detailed in Appendix C.2.6.

Therefore, the reported improvements are not based on a single experimental run, and their variability is already clearly communicated in our experimental results.

4. Normalization and zero reference values

We appreciate the opportunity to clarify this point. The numerical stability issue raised by the reviewer is already explicitly addressed in the appendix.

Appendix B.1 introduces a positive stabilization constant in the normalization denominator. Furthermore, Appendix C.1 provides the exact normalization used in the experiments, including denominator floors of 10-8 for synthetic experiments and 0.05 for KuaiRec.

Therefore, the implementation does not suffer from the undefined division-by-zero issue suggested by the main-text expression. The normalization is well-defined under the experimental protocols, and the relevant implementation details are already documented.

We acknowledge that the main-text notation can be made more consistent with the appendix, and we will clarify this presentation detail. However, this is a notational clarification rather than a flaw in the implemented method or the reported experimental results.

5. Relationship to user satisfaction and churn

We thank the reviewer for this comment. However, we would like to clarify that our paper does not claim that ODR has been causally demonstrated to reduce user churn.

Figure 1 presents an observational analysis motivating the importance of avoiding severe recommendation dissatisfaction. The distinction between motivation and causal evidence is already explicitly stated in Sections 6 and 7, which notes that demonstrating improvements in retention or subscription revenue would require a production A/B test.

Our central contribution is a robust recommendation objective that protects against poor performance on plausible user objectives, not a causal estimation method for user retention. Requiring a demonstration of causal churn reduction would therefore introduce an additional research question beyond the scope of our stated contribution. We believe the current motivation and empirical evaluation are appropriate for the objective studied in this paper.

We would like to kindly remind the reviewer that we have submitted a revised manuscript, with all changes highlighted in red, addressing the key concerns raised in the review. In particular, we have clarified the optimization mechanism and learning dynamics, statistical variability across independent runs, numerical stabilization, and the scope of our real-world evaluation.

We would greatly appreciate it if the reviewer could review these changes and identify any remaining substantive concerns before the discussion deadline. In the absence of further objections, we believe the key questions raised in the review have been satisfactorily resolved, and we hope that the reviewer will take these revisions into account when finalizing their assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback and for recognizing the importance of robust recommendation under uncertain and varying user intents.

We would like to emphasize that the central contribution of ODR is a principled formulation of recommendation under unknown, potentially time-varying user intent distributions, rather than a new general-purpose optimization primitive. Specifically, we formulate per-context worst-objective optimization, establish its minimax interpretation and gradient characterization, and empirically demonstrate its effectiveness across synthetic and real-world recommendation settings.

We believe that several concerns raised by the reviewers are already addressed in the main text or appendices, while others concern additional guarantees or experimental settings beyond our stated contributions.

Below, we clarify these points and address the reviewers' questions. We maintain that the identified presentation issues and opportunities for further investigation do not invalidate our central formulation or the empirical findings supporting it.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.