Learning to Fast Unrank in Collaborative Filtering Recommendation

TL;DR

This work tackles privacy-preserving requirements in recommender systems by reframing unlearning as unranking, where target items are deliberately demoted in ranking rather than removed wholesale. It introduces L2UnRank, a model-agnostic framework built on three components: localized interaction-based influence scoping, fine-grained influence quantification (combining structural and semantic cues), and a weighted influence function implemented via a conjugate gradient-based update of a ranking-oriented loss. The approach achieves state-of-the-art unranking effectiveness with substantial speedups over retraining and other baselines, while preserving recommendation utility across multiple datasets and backbone models. The results demonstrate practical viability for real-time privacy requests and provide a principled, scalable path toward privacy-aware recommendation systems.

Abstract

Modern data-driven recommendation systems risk memorizing sensitive user behavioral patterns, raising privacy concerns. Existing recommendation unlearning methods, while capable of removing target data influence, suffer from inefficient unlearning speed and degraded performance, failing to meet real-time unlearning demands. Considering the ranking-oriented nature of recommendation systems, we present unranking, the process of reducing the ranking positions of target items while ensuring the formal guarantees of recommendation unlearning. To achieve efficient unranking, we propose Learning to Fast Unrank in Collaborative Filtering Recommendation (L2UnRank), which operates through three key stages: (a) identifying the influenced scope via interaction-based p-hop propagation, (b) computing structural and semantic influences for entities within this scope, and (c) performing efficient, ranking-aware parameter updates guided by influence information. Extensive experiments across multiple datasets and backbone models demonstrate L2UnRank's model-agnostic nature, achieving state-of-the-art unranking effectiveness and maintaining recommendation quality comparable to retraining, while also delivering a 50x speedup over existing methods. Codes are available at https://github.com/Juniper42/L2UnRank.

Figures (12)

• Figure 1: An example of unranking. The red box represents the items to be forgotten.
• Figure 2: Illustration of recommendation unlearning scenarios. 0-hop items and users refer to directly influenced entities, they and the interactions between them are namely the 0-hop neighborhood.
• Figure 3: Overview of the proposed L2UnRank method. Deleting an item is taken as an example.
• Figure 4: Effect of influenced scope size $p$ after randomly deleting 5% of items. The line represents URR, and the bar represents Time (s).
• Figure 5: Robustness evaluation under different Interaction Unlearning ratios.