acceptodds
Under review as a conference paper at ICLR 2027

SREraser: Efficient Exact Unlearning for Sequential Recommendation

Abstract

Sequential recommender systems ingest sensitive personal histories, necessitating machine unlearning to ensure privacy compliance and maintain system utility. Since retraining from scratch is costly and approximate unlearning cannot certify complete data removal, exact unlearning via data sharding has become the prevailing solution. However, most studies in recommendation unlearning have focused on collaborative filtering, and the application of this paradigm to sequential recommendation remains underexplored. Existing methods face two limitations: partitioning strategies scale poorly with the number of users, and aggregators assign a single scalar weight to each sub-model, ignoring how sub-model reliability varies across representation dimensions. In this paper, we propose SREraser, a framework capable of interaction-level unlearning in sequential recommendation. SREraser introduces two core designs. First, User-Centric Spherical Partitioning (UCSP) ranks shards for each user by cosine similarity and resolves shard capacity constraints locally, reducing partitioning cost by an order of magnitude. Second, Adaptive Dimension Aggregation (ADA) produces one weight per representation dimension of every sub-model, conditioned on how closely that sub-model's response aligns with its shard prototype. Experiments on three real-world datasets show that SREraser matches or outperforms other shard-based unlearning methods in recommendation utility, while maintaining comparable unlearning efficiency.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
machine unlearning, sequential recommendation, exact unlearning, data partitioning, model aggregation, right to be forgotten

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_Ww19OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SREraser addresses interaction-level unlearning in sequential recommendation through two components: User-Centric Spherical Partitioning (UCSP), which reduces user-partitioning overhead, and Adaptive Dimension Aggregation (ADA), which combines sub-model representations using dimension-wise weights. Experiments demonstrate improved recommendation utility over the evaluated shard-based baselines, with comparable unlearning time.

Critical strengths

S1. Effective partitioning optimization. UCSP substantially reduces partitioning overhead while maintaining comparable recommendation quality. On Games, partitioning takes 1.71 seconds, compared with 8.06 seconds for spherical KMeans and 28.20 seconds for Euclidean KMeans.

Critical weaknesses

W1. The central exact-unlearning guarantee is not established. UCSP constructs shards using an encoder pretrained on the full training dataset, so shard assignments can retain influence from subsequently deleted interactions. Retraining only the affected sub-model and aggregator does not establish equivalence to rerunning the complete pipeline on retained data. The authors need to account for this data-dependent partitioning in their exactness argument.

W2. The efficiency evaluation does not match the utility workload. Utility is evaluated after deletions distributed across every shard, whereas timing covers a single interaction-deletion request. Moreover, cache-refresh costs are partly estimated using a theoretical scaling factor. These results therefore do not establish end-to-end efficiency for the batch-deletion workload used to evaluate utility.

W3. Novelty is incremental, and ADA’s distinctive mechanism is insufficiently isolated. SREraser retains SRU’s shard-based framework and primarily modifies partitioning and aggregation. Although ADA improves accuracy, the ablations lack a prototype-free dimension-wise gate and capacity-matched alternatives. Consequently, they do not establish whether prototype conditioning specifically contributes beyond finer-grained weighting or increased model capacity.

Questions for authors

Please refer to Weaknesses.

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

This paper proposes SREraser, a framework for data unlearning in sequential recommendation. The method introduces two components. First, User-Centric Spherical Partitioning improves efficiency by replacing global sorting with randomized user ordering and local shard selection. Second, Adaptive Dimension Aggregation generates dimension-level weights to combine representations from different sub-models. Experiments on three datasets and three sequential recommendation backbones show that SREraser improves recommendation performance over existing unlearning methods while reducing computational cost.

Critical strengths
  1. Both proposed components have reasonable motivations. UCSP reduces the computational cost while ADA provides more flexible aggregation.
  2. The paper evaluates the proposed framework on three datasets and three sequential recommendation backbones, which demonstrate the practical utility of the framework.
Critical weaknesses
  1. Limited methodological novelty and weak task-specific motivation. The two proposed components mainly address general problems in clustering and representation aggregation rather than challenges specific to sequential recommendation unlearning. UCSP improves the efficiency of capacity-constrained clustering through a simple modification on the assignment strategy, which could be applied to any general clustering tasks. Similarly, ADA introduces dimension-wise weights for representation aggregation, an idea already explored in existing studies on expert fusion. Therefore, the methodological novelty of both components appears limited, and their necessity for sequential recommendation unlearning is insufficiently justified. Moreover, the two components lack a clear methodological connection: UCSP focuses on reducing computational costs, whereas ADA introduces a more complex aggregation mechanism. The paper does not clearly explain how these two designs jointly address a specific challenge in recommendation unlearning.
  2. Limited practical significance and insufficient convergence analysis of UCSP. UCSP improves partitioning efficiency on initialization rather than the unlearning step. Table 3 reports comparable unlearning times for SREraser and SRU.Intuitively, initialization is a infrequent operation, whereas unlearning may occur repeatedly, making the cost of processing unlearning requests more important in practice. Therefore, the practical significance of this optimization for the overall unlearning process remains unclear. Moreover, I am concerned that UCSP may converge more slowly than existing methods. Since UCSP randomly permutes users in each iteration, different user orders may produce different assignments under capacity constraints, potentially requiring more iterations to converge and offsetting its efficiency advantage. More importantly, although Algorithm 1 specifies convergence-based early stopping, the released code appears to use a fixed number of iterations (T=50) without implementing this stopping condition. This raises concerns about the fairness of the efficiency comparison, as different methods may require different numbers of iterations to converge.
Questions for authors
  1. UCSP aims to improve computational efficiency, while ADA introduces a more complex aggregation mechanism. How do the authors justify combining these two designs from a practical deployment perspective?
  2. What stopping strategies were used for UCSP, KMeans, and SKMeans in the reported experiments? Could the authors provide their iteration counts and runtime comparisons under same stopping conditions?
  3. How does ADA differ from existing dimension-wise expert weighting or fine-grained attention mechanisms? (e.g. "A Multitask Learning Model with Multiperspective Attention and Its Application in Recommendation")
1 comment

Thank you for your detailed response.

I still have concerns about the motivation and methodological novelty. Essentially, UCSP modifies the capacity-constrained K-means by introducing randomized user ordering, while ADA extends the aggregation gate to dimension-wise weighting. These still appear to be two relatively simple. For example, replacing UCSP with another efficient K-means variant, or ADA with another aggregation mechanism, might also lead to some improvements. I still do not see a sufficiently strong connection between these designs and the unique challenges of machine unlearning, or why their combination constitutes a substantial methodological contribution to this field.

I appreciate the clarification regarding the efficiency comparison, but my convergence concern remains. Running all methods for a fixed 50 iterations ensures the same iteration budget, but does not necessarily provide a fair comparison of their efficiency in reaching convergence. For example, what if conventional K-means converges within 20 iterations, while UCSP requires 50 due to its randomized assignment strategy? I would therefore like to see a comparison based on actual convergence behavior rather than only a fixed iteration budget.

For these reasons, I will maintain my original rating.

Reviewer_nrr3OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SREraser, an unlearning framework for sequential recommendation. It has two main components: User-Centric Spherical Partitioning (UCSP) and Adaptive Dimension Aggregation (ADA). UCSP assigns users to shards based on their sequence embeddings, while ADA combines the outputs of different shard models using dimension-wise weights. When an interaction needs to be deleted, the affected shard model is retrained and the aggregator is trained again. Experiments on three datasets and three recommendation backbones show improved Recall@20 compared to the considered unlearning baselines. However, the exact unlearning guarantee and the claimed efficiency improvement need further validation.

Critical strengths
  1. The ADA aggregator gives consistent utility improvements over existing aggregation methods, especially on the sparse Games dataset. The improvements are observed with all three recommendation backbones. This is a useful contribution to shard-based recommendation unlearning.

  2. The paper studies a practical problem: deleting individual interactions from sequential recommender systems. The evaluation covers three datasets and three different backbones, which gives a reasonable view of the method's performance across different settings.

  3. The proposed framework combines shard-based retraining with adaptive aggregation. The reported results show that it can improve recommendation utility while being faster than retraining the full model from scratch.

Critical weaknesses
  1. The exact unlearning claim is not established. UCSP uses a SASRec encoder pre-trained on the full dataset, including the interactions that may later be deleted. The partition is then kept fixed during unlearning. Hence, the partition itself can still depend on the forget set. The paper needs to explain how this satisfies its definition of exact unlearning.

  2. The efficiency claim is not sufficiently supported. The partitioning step is a one-time cost and does not appear to reduce the actual unlearning time. In Table 3, SREraser and SRU have almost the same unlearning time on Games with SASRec. The paper should report measured end-to-end times, including cache updates, and clarify which reported values are estimates.

  3. The source of ADA's improvement is not isolated clearly. The current ablations do not separate the benefit of dimension-wise weighting from input adaptivity, prototype agreement, and the additional model parameters. Important comparisons with existing recommendation-unlearning methods are also missing.

Questions for authors
  1. The partitioning encoder is pre-trained on the full dataset. Is it retrained, or is the partition recomputed after an interaction is deleted? If not, how does the method satisfy Definition 3.2 of exact unlearning?

  2. Which timings in Table 3 are measured and which are estimated using Eq. 10? Can the authors report measured end-to-end unlearning time, including cache updates, with standard deviations?

  3. Can the authors compare ADA against input-adaptive scalar weights with a similar parameter count, static dimension-wise weights, and a version without the prototype-agreement input? These comparisons would help establish which part of ADA provides the improvement.

  4. What are the paired units used in the reported statistical tests? Can the authors report standard deviations for Table 2 and clarify the significance of small differences between SREraser and SRU?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.