PaperlayerView exact revision on arXiv

Source-linked AI summary

Recommendation Unlearning

Chong Chen, Fei Sun, Min Zhang, Bolin Ding

arXiv:2201.06820v2cs.IRcs.LG

TL;DR

Recommendation systems may need to erase the influence of sensitive or harmful training data, but existing unlearning methods do not directly preserve recommendation-specific collaborative information. RecEraser uses balanced similarity-based sharding and adaptive aggregation for recommendation unlearning. Experiments on three real-world datasets and three recommendation models report efficient unlearning and higher model utility than state-of-the-art unlearning methods.

  • Problem

    Existing machine-unlearning methods cannot be directly applied to recommendation because random sharding can destroy collaborative information, while recommendation systems may need to forget specified training data.

  • Method

    RecEraser partitions recommendation training data into balanced similarity-based shards, trains submodels, and uses adaptive aggregation to combine their predictions.

  • Results

    Across three real-world datasets and three representative recommendation models, RecEraser achieves efficient unlearning and outperforms state-of-the-art unlearning methods in model utility.

  • Takeaways & Limitations

    RecEraser provides a recommendation-specific framework designed to combine efficient unlearning with preserved collaborative information and model utility.

  • Takeaways & Limitations

    The paper focuses on sequential unlearning requests; efficient batch unlearning with SISA remains an open problem, and future work includes content features and related tasks.

Abstract

from arXiv · show

Recommender systems provide essential web services by learning users' personal preferences from collected data. However, in many cases, systems also need to forget some training data. From the perspective of privacy, several privacy regulations have recently been proposed, requiring systems to eliminate any impact of the data whose owner requests to forget. From the perspective of utility, if a system's utility is damaged by some bad data, the system needs to forget these data to regain utility. From the perspective of usability, users can delete noise and incorrect entries so that a system can provide more useful recommendations. While unlearning is very important, it has not been well-considered in existing recommender systems. Although there are some researches have studied the problem of machine unlearning in the domains of image and text data, existing methods can not been directly applied to recommendation as they are unable to consider the collaborative information. In this paper, we propose RecEraser, a general and efficient machine unlearning framework tailored to recommendation task. The main idea of RecEraser is to partition the training set into multiple shards and train a constituent model for each shard. Specifically, to keep the collaborative information of the data, we first design three novel data partition algorithms to divide training data into balanced groups based on their similarity. Then, considering that different shard models do not uniformly contribute to the final prediction, we further propose an adaptive aggregation method to improve the global model utility. Experimental results on three public benchmarks show that RecEraser can not only achieve efficient unlearning, but also outperform the state-of-the-art unlearning methods in terms of model utility. The source code can be found at https://github.com/chenchongthu/Recommendation-Unlearning

1 INTRODUCTION

Recommendation unlearning addresses the need to remove specified training-data influence while preserving recommendation quality. RecEraser partitions data into balanced similarity-based shards and adaptively aggregates shard models, achieving efficient unlearning and stronger recommendation performance than prior unlearning methods.

  • Motivation: RecEraser targets recommendation unlearning, where sensitive or harmful training data must be forgotten without discarding the collaborative information needed for recommendations.Existing unlearning methods from vision and language cannot directly account for recommendation-specific user-item collaboration.
  • Challenges: Random sharding can damage recommendation performance because user-item interactions contain collaborative information across users and items.Existing aggregation schemes may also assign static or non-adaptive weights to shard models.
  • Method: RecEraser divides training data into multiple shards and trains a constituent submodel for each shard to support efficient unlearning.Only affected shard models need to be retrained under the shard-based design.
  • Method: Three balanced partition strategies use user, item, and interaction similarities to preserve collaborative information without letting shard-size imbalance reduce unlearning efficiency.The strategies are designed specifically for balanced recommendation-data partitions.
  • Method: An attention-based adaptive aggregation method accounts for different shard contributions when predicting different user-item pairs.This design is intended to improve the global recommendation utility beyond static aggregation.
  • Evaluation: Across three real-world datasets and three representative recommendation models, RecEraser achieves efficient unlearning and outperforms the state-of-the-art unlearning framework in recommendation performance.The reported experiments evaluate both unlearning efficiency and recommendation utility.

2 RELATED WORK

Prior recommendation methods learn from explicit or implicit user feedback using factorization, sampling, and neural architectures. Machine unlearning removes specified training-data influence through approximate or exact approaches, but existing methods face efficiency, guarantee, and recommendation-specific collaboration challenges.

  • Recommendation Methods: Recommendation research progressed from latent-factor models for explicit feedback to methods for implicit feedback, including WMF and pairwise BPR learning.WMF treats unobserved items as negative samples, whereas BPR optimizes relative preferences over item pairs.
  • Recommendation Methods: Neural Collaborative Filtering jointly learns matrix factorization and a feedforward neural network, motivating broader use of neural architectures and attention mechanisms in recommendation.The cited literature extends neural recommendation to different scenarios and architectures.
  • Machine Unlearning: Machine unlearning removes the influence of a specified subset of training data from a trained model and is commonly divided into approximate and exact unlearning.Approximate unlearning provides statistical rather than complete deletion guarantees.
  • Machine Unlearning: Approximate unlearning is more efficient but offers probabilistic guarantees that are difficult to apply to non-convex models and may be unsuitable for regulated recommender systems.These methods often use gradient-based updates or influence-function approaches to remove requested samples quickly.
  • Machine Unlearning: Exact unlearning seeks complete removal of requested data and SISA accelerates it by independently training models on disjoint shards and retraining only affected submodels.SISA aggregates the independently trained shard results for final prediction.
  • Recommendation Unlearning: Recommendation unlearning remains challenging because random partitioning destroys collaborative information, while clustering-based partitioning can produce highly unbalanced shards.These trade-offs affect recommendation performance and unlearning efficiency, respectively.

3 RECOMMENDATION UNLEARNING

Recommendation unlearning requires removing a revoked interaction’s influence while preserving efficient retraining and comparable recommendation performance. Existing unlearning methods do not directly fit recommendation because random sharding can destroy collaborative information.

  • Recommendation unlearning seeks a model trained on Y\y_uv after a user revokes interaction y_uv.
  • The task requires provable removal of revoked data, high unlearning efficiency, and performance comparable to retraining from scratch.
  • Existing machine-unlearning methods cannot be directly applied because recommendation inputs contain collaborative user-item information.

4 RECERASER METHOD

RecEraser partitions recommendation data into balanced shards, trains a submodel per shard, and adaptively aggregates their representations for prediction. Upon an unlearning request, only the affected submodel and aggregation component require retraining.

  • 4.1 Overview: RecEraser uses balanced data partition, submodel training, and attention-based adaptive aggregation to preserve collaborative information.
  • 4.1 Overview: Only the submodel containing the revoked point and the aggregation component are retrained, rather than the whole model.
  • 4.2 Balanced Data Partition: UBP and IBP partition by user or item similarity, while InBP uses user-item interaction similarity to capture global interaction features.
  • 4.3 Attention-based Adaptive Aggregation: Each shard trains a submodel, and attention weights vary by prediction because different shards contribute differently to user-item pairs.
  • 4.3 Attention-based Adaptive Aggregation: The aggregation first maps submodel embeddings into a shared representation space, then combines them using learned user and item attention weights.
  • 4.3 Attention-based Adaptive Aggregation: The aggregation is retrained after any request involving the full training data and takes about 10 epochs, contributing to unlearning time.
  • 4.4 Training: RecEraser is model-agnostic and is instantiated with BPR, WMF, and LightGCN recommendation models.

5.1 Experimental Setup

The experiments evaluate unlearning methods on three public recommendation datasets using representative recommendation models and ranking metrics. Data are converted to implicit feedback and assessed with Recall@N and NDCG@N.

  • Experiments use Yelp2018, Movielens-1m, and Movielens-10m after converting ratings into binary implicit feedback.
  • Comparisons include Retrain, SISA, and GraphEraser alongside RecEraser across different recommendation models.
  • Each dataset uses 80% of interactions for training, with 10% of the training interactions held out for validation.
  • Recommendation quality is evaluated with Recall@N and position-aware NDCG@N, using N values of 10, 20, and 50.

5.2 Recommendation Performance

RecEraser achieves significantly better recommendation performance than state-of-the-art unlearning baselines across three datasets, while retraining generally remains strongest but inefficient.

  • Retraining generally achieves the best recommendation performance, but its unlearning efficiency is low for complex models and large datasets.
  • RecEraser significantly outperforms state-of-the-art unlearning baselines on all three datasets (p<0.01).
  • RecEraser improves over SISA on Yelp2018 by 59%, 41%, and 46% with BPR, WMF, and LightGCN, respectively.
  • RecEraser performs more comparably to retraining on the denser Movielens-1m and Movielens-10m datasets than on sparse Yelp2018.

5.3 Unlearning Efficiency

RecEraser substantially reduces unlearning time because only affected shard models and the aggregation component require retraining.

  • On Movielens-10m, RecEraser requires 320, 16, and 920 minutes for optimal BPR, WMF, and LightGCN performance, versus 3250, 292, and 25330 minutes for retraining.
  • RecEraser achieves more than 10× acceleration over retraining on Movielens-10m.
  • RecEraser improves efficiency by retraining only the corresponding submodel and aggregation component after an unlearning request.

5.4 Ablation Study

The ablations show that similarity-based partitioning and attention-based aggregation both improve recommendation performance over their respective baselines.

  • 5.4.1 Effect of proposed data partition methods.: UBP, IBP, and InBP outperform Random partitioning on Yelp2018 and Movielens-1m.
  • 5.4.1 Effect of proposed data partition methods.: InBP reaches Recall@20 of 0.0568 versus 0.0472 for Random on Yelp2018 with LightGCN.
  • 5.4.1 Effect of proposed data partition methods.: InBP generally performs better than UBP and IBP, indicating that interaction-based partitioning is more effective than user- or item-only partitioning.
  • 5.4.2 Effect of proposed attention-based aggregation method.: On Yelp2018 with WMF, AttAgg achieves Recall@20 of 0.0572 versus 0.0475 for MeanAgg and 0.0482 for StaticAgg.
  • 5.4.2 Effect of proposed attention-based aggregation method.: AttAgg adaptively assigns different submodel weights for different user-item predictions, unlike fixed averaging or static weighting.

5.5 Impact of the Shard Number

Increasing the shard count reduces unlearning time but slightly decreases recommendation performance, creating a trade-off that requires careful shard selection.

  • Average unlearning time decreases as the number of shards increases across all recommendation models.
  • Figure 4 jointly reports recommendation performance and unlearning time cost for different shard counts on Yelp2018 and Movielens-1m.
  • Increasing shard count slightly decreases recommendation performance because collaborative information generally declines across smaller shards.
  • The shard number should be selected to balance unlearning efficiency against recommendation performance.

6 CONCLUSION AND FUTURE WORK

RecEraser is presented as a recommendation-specific framework that combines efficient unlearning with high recommendation performance. The work focuses on sequential unlearning requests, while batch unlearning and incorporating content features remain future directions.

  • RecEraser achieves efficient unlearning and outperforms state-of-the-art unlearning methods in model utility.The evaluation covers three real-world datasets and three representative recommendation models.
  • The framework is designed to preserve collaborative information while supporting erasable recommendation models.The paper identifies erasable models as increasingly important as data-protection requirements become prevalent in real applications.
  • The study evaluates sequential unlearning requests, leaving efficient batch unlearning as an open problem for future work.Future directions also include incorporating user and item content features and exploring related tasks such as social network embedding.
Loading 2201.06820v2…