Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Making Recommender Systems Forget: Learning and Unlearning for Erasable Recommendation

Yuyuan Li, Xiaolin Zheng, Chaochao Chen, Junlin Liu

TL;DR

This paper tackles privacy-driven unlearning in recommender systems by introducing LASER, a framework with two modules that preserve collaborative information while enabling erasure. LASER uses a hypergraph-based approach to learn collaborative embeddings, forms balanced user groups, and trains groups in an easy-to-hard sequence (SeqTrain) to enhance model utility. The authors provide theoretical analysis showing benefits of the SeqTrain order and validate the approach with experiments on MovieLens 1M and Amazon Digital Music, demonstrating efficient unlearning and improved utility over state-of-the-art baselines. The work advances practical, scalable unlearning for CF models, enabling compliant, privacy-preserving recommendations without sacrificing performance.

Abstract

Privacy laws and regulations enforce data-driven systems, e.g., recommender systems, to erase the data that concern individuals. As machine learning models potentially memorize the training data, data erasure should also unlearn the data lineage in models, which raises increasing interest in the problem of Machine Unlearning (MU). However, existing MU methods cannot be directly applied into recommendation. The basic idea of most recommender systems is collaborative filtering, but existing MU methods ignore the collaborative information across users and items. In this paper, we propose a general erasable recommendation framework, namely LASER, which consists of Group module and SeqTrain module. Firstly, Group module partitions users into balanced groups based on their similarity of collaborative embedding learned via hypergraph. Then SeqTrain module trains the model sequentially on all groups with curriculum learning. Both theoretical analysis and experiments on two real-world datasets demonstrate that LASER can not only achieve efficient unlearning, but also outperform the state-of-the-art unlearning framework in terms of model utility.

Making Recommender Systems Forget: Learning and Unlearning for Erasable Recommendation

TL;DR

This paper tackles privacy-driven unlearning in recommender systems by introducing LASER, a framework with two modules that preserve collaborative information while enabling erasure. LASER uses a hypergraph-based approach to learn collaborative embeddings, forms balanced user groups, and trains groups in an easy-to-hard sequence (SeqTrain) to enhance model utility. The authors provide theoretical analysis showing benefits of the SeqTrain order and validate the approach with experiments on MovieLens 1M and Amazon Digital Music, demonstrating efficient unlearning and improved utility over state-of-the-art baselines. The work advances practical, scalable unlearning for CF models, enabling compliant, privacy-preserving recommendations without sacrificing performance.

Abstract

Privacy laws and regulations enforce data-driven systems, e.g., recommender systems, to erase the data that concern individuals. As machine learning models potentially memorize the training data, data erasure should also unlearn the data lineage in models, which raises increasing interest in the problem of Machine Unlearning (MU). However, existing MU methods cannot be directly applied into recommendation. The basic idea of most recommender systems is collaborative filtering, but existing MU methods ignore the collaborative information across users and items. In this paper, we propose a general erasable recommendation framework, namely LASER, which consists of Group module and SeqTrain module. Firstly, Group module partitions users into balanced groups based on their similarity of collaborative embedding learned via hypergraph. Then SeqTrain module trains the model sequentially on all groups with curriculum learning. Both theoretical analysis and experiments on two real-world datasets demonstrate that LASER can not only achieve efficient unlearning, but also outperform the state-of-the-art unlearning framework in terms of model utility.
Paper Structure (49 sections, 2 theorems, 15 equations, 13 figures, 1 table, 4 algorithms)

This paper contains 49 sections, 2 theorems, 15 equations, 13 figures, 1 table, 4 algorithms.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Unlearning
  4. Goals of Unlearning
  5. Choice of Recommendation Models
  6. LASER Framework
  7. Overview
  8. Group Module
  9. Collaboration Principle
  10. Motivation
  11. Method
  12. Balance Principle
  13. Motivation
  14. Method
  15. SeqTrain Module
  16. ...and 34 more sections

Key Result

Theorem 1

Given a Bayesian prior $p$ for the model parameters $\theta$, we have:

Figures (13)

  • Figure 1: A schematic view of learning and unlearning in recommendation.
  • Figure 2: Overview of LASER framework.
  • Figure 3: An illustration of an user-item bipartite graph and the high-order relations of $u_1$ where $u$ and $i$ denote user and item respectively.
  • Figure 4: Comparison results of running time.
  • Figure 5: NDCG@10 and HR@10 results for utility (G3) study and ablation study (Group module) on ML dataset with std under 1e-3.
  • ...and 8 more figures

Theorems & Definitions (6)

  • Theorem 1
  • proof
  • Definition 1: User's $l$-order reachable neighbors
  • proof
  • Theorem 2
  • proof