acceptodds
Under review as a conference paper at ICLR 2027

RecDataAgent: User-Conditioned Data Augmentation for Sequential Recommendation

Abstract

Improving sequential recommendation requires not only learning from user behavior, but also deciding how that behavior becomes training supervision. With heterogeneous preferences, the value of an augmented view depends on whose history it represents and which signals it preserves. We therefore frame data augmentation as the optimization of supervision construction: user selection, method assignment, and exposure form a complete policy whose value is determined by the recommendation outcomes of the resulting corpus. We introduce RecDataAgent, an agent that connects user understanding, executable data construction, and outcome-guided revision. It analyzes train-visible user context to propose augmentation policies, materializes them while preserving task targets and the construction budget, and retains each policy together with its execution trace and measured outcomes. This feedback allows both user partitions and method assignments to evolve, making data construction a decision process informed by the learning it supports. Experiments across sequential recommendation domains show that contextual assignment improves retrieval and ranking under matched method exposure, while outcome feedback strengthens policies beyond user analysis and execution checks. These findings support a shift from designing augmentation methods in isolation to optimizing how they jointly shape supervision for different users. RecDataAgent makes this shift concrete through a closed loop in which user context guides construction and recommendation outcomes guide its revision. Anonymous code is available at https://anonymous.4open.science/r/RecDataAgent-382D.

Primary area:
generative models
Keywords:
Sequential Recommendation, Data Agent

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_t4fuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

For LLM-based sequential recommendation, an LLM planner routes users via 17 hand-built features to existing augmentation views (L2Aug, GenPAS prefix, AsarRec dual) under a fixed record budget. Candidates are scored by one-epoch Qwen3-4B fine-tuning. The validation-best of nine (three revision rounds) beats the best single method on Amazon Beauty, Toys, Sports (R@10 +4.59%, +12.35%, +18.30%), fixed-total routings, and a no-feedback planner.

Critical strengths

Matched-exposure controls (Table 3) fix method totals while shuffling who gets which view, isolating user routing from method mix. Toys and Sports R@10 gains are sizable (0.0828 vs. 0.0771, 0.0569 vs. 0.0480).

The policy is an inspectable rule list (Beauty: cluster_id==1 routes to GenPAS prefix, else AsarRec dual, p. 19), and the planner sees only aggregate validation metrics (p. 14).

Critical weaknesses

The LLM agent is not shown to beat picking the validation-best of nine candidates (pp. 6, 14), since baselines are single runs without budget-matched non-LLM search. Without feedback, the planner trails uniform AsarRec (Beauty 0.0770 vs. 0.0784, Sports 0.0456 vs. 0.0481, Tables 2–4), and displayed revisions do not improve test scores even when Sports validation rises (Tables 5–6), so gains may come from nine-try selection.

There are no SDs or significance tests, and Tables 3–5 omit seed counts. The same Beauty policy scores R@10 0.0800 (Tables 6–7) and 0.0820 (Tables 2, 5), so NDCG claims on p. 9 and p. 15 conflict. Beauty controls (0.0685–0.0717) also trail both uniform methods (0.0730, 0.0784), suggesting variance above the +0.0036 margin or a protocol difference.

Minor:

Every user gets one appended record (p. 6), so the policy reduces to method assignment. Missing: no-augmentation row, no-transfer control, and compute cost (pp. 6, 8). Related work omits close prior work: AdaAug (2022), LLM-AutoDA (2024), and DataEnvGym (2025).

Questions for authors

How does non-LLM search with the same nine-candidate budget (e.g., enumerating cluster-to-method assignments, random routings) compare?

What are per-seed standard deviations? Why is the Beauty Ref 0.0800 in Tables 6–7 but 0.0820 in Tables 2 and 5?

Reviewer_yt5YOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RecDataAgent, an agent-based framework that establishes a closed loop for sequential recommendation through policy proposal, execution, outcome analysis and summarization, policy updating, archiving, and revision suggestions. Rather than applying a single data augmentation method uniformly, the framework optimizes which users receive which augmented views and how the augmentation budget is allocated. The motivation is that users have heterogeneous interaction histories and behavioral characteristics, meaning that different augmentation methods may provide different benefits to different users. Experiments on three Amazon datasets—Beauty, Toys, and Sports—show that the proposed framework improves recommendation performance over several existing augmentation methods and alternative routing strategies under matched augmentation budgets. The results suggest that user-conditioned assignment and outcome-guided policy revision can improve the effectiveness of data augmentation.

Critical strengths
  1. The paper addresses an important problem by integrating user-conditioned assignment, executable policy construction, and outcome-guided revision into a unified framework. Its main contribution lies in combining and operationalizing these components for sequential recommendation, rather than proposing an entirely new augmentation operator.
  2. The framework connects policy selection, method combination, execution, evaluation, reflection, archiving, and reuse. Its modular design, explicit execution constraints, and provenance tracking make it potentially useful for integration into existing recommendation systems.
  3. The Beauty case study illustrates how behavioral characteristics, including category concentration, category switching, and item-description overlap, can inform augmentation routing decisions. This provides some interpretability regarding why different users may be assigned different augmentation methods.
Critical weaknesses
  1. Although the framework provides a mechanism for assigning different methods to different users, it does not establish clear principles for selecting appropriate methods from the numerous existing data augmentation techniques. The method catalog and policy-search process are central to the framework, but the criteria for choosing, combining, or introducing operators remain largely heuristic. More systematic guidance would improve the framework's generality and practical usability.
  2. The evaluation covers only three Amazon product categories that share the same data source and general experimental setup. The authors should evaluate a broader range of conventional next-item recommendation tasks, datasets, and recommendation architectures to verify whether the proposed framework generalizes beyond the current setting.
  3. Although the main local comparisons use five training seeds, the paper primarily reports mean performance without standard deviations, confidence intervals, or statistical significance tests. Consequently, it is difficult to determine whether the observed improvements are statistically reliable, particularly when the performance gains are relatively small.
  4. The framework evaluates only nine candidate policies per domain, with three candidates per round over three rounds. However, it does not compare the proposed controller against equally budgeted random search, evolutionary search, Bayesian optimization, or other systematic policy-search methods. Therefore, it remains unclear whether the improvements arise specifically from the agent's reasoning and outcome-feedback mechanism or simply from exploring alternative augmentation configurations.
Questions for authors
  1. How does the framework identify weaknesses in the current policy, and how are these weaknesses incorporated into the next iteration? The paper describes how recommendation outcomes and execution diagnostics guide policy revision, but it is unclear how the framework diagnoses the underlying causes of poor performance and translates them into actionable revision suggestions.
  2. How does the framework localize and address failures when multiple augmentation strategies and routing rules are combined? When a policy performs poorly, how can the authors distinguish whether the failure originates from user partitioning, method selection, exposure allocation, or the execution of an individual augmentation operator? Can the framework identify the problematic component and revise it without unnecessarily changing the other components?
  3. Could the authors clarify the statement in Line 309 that improvements occur in “all twelve domain–metric comparisons”?
  4. Why do different augmentation baselines perform best on different datasets, and how does RecDataAgent account for these differences?
Reviewer_7cs5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how to assign different augmentation methods personally. RecDataAgent proposes method from user features and revises them using validation results. The selected policies outperform locally evaluated augmentation baselines on public datasets.

Critical strengths

User routing improves performance when method exposure is fixed. This control supports the value of choosing which method each user receives. The feedback ablation further supports using validation outcomes to improve policies.

Critical weaknesses

Without a non-LLM baseline under the same budget, the contribution remains unclear.

Questions for authors

Was non-LLM search method tested with the same nine-evaluation budget?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.