Privacy-Preserving Cross-Domain Sequential Recommendation

TL;DR

This work tackles privacy in cross-domain sequential recommendation by introducing sequential differential privacy (SDP) and a non-invasive two-stage framework, PriCDSR, that perturbs auxiliary-domain sequences with a novel random mechanism while leaving the base CDSR model unchanged. The mechanism $ ext{M}$ guarantees $oldsymbol{ ext{$oldsymbol{ ext{$ ext{epsilon}$}}}$-SDP$, protecting both item IDs and sequence order, and its perturbation yields $ ilde{oldsymbol{R}}^A$ usable by any SFCDSR algorithm in the target domain. Empirical results on six Amazon domain pairs show PriCDSR maintains competitive recommendation performance relative to single-domain baselines and underlines the privacy-utility trade-off governed by the privacy budget $oldsymbol{ ext{epsilon}}$. Overall, the paper provides a principled privacy framework for CDSR that can be adopted without modifying existing base models, enabling legally and ethically responsible cross-domain data utilization.

Abstract

Cross-domain sequential recommendation is an important development direction of recommender systems. It combines the characteristics of sequential recommender systems and cross-domain recommender systems, which can capture the dynamic preferences of users and alleviate the problem of cold-start users. However, in recent years, people pay more and more attention to their privacy. They do not want other people to know what they just bought, what videos they just watched, and where they just came from. How to protect the users' privacy has become an urgent problem to be solved. In this paper, we propose a novel privacy-preserving cross-domain sequential recommender system (PriCDSR), which can provide users with recommendation services while preserving their privacy at the same time. Specifically, we define a new differential privacy on the data, taking into account both the ID information and the order information. Then, we design a random mechanism that satisfies this differential privacy and provide its theoretical proof. Our PriCDSR is a non-invasive method that can adopt any cross-domain sequential recommender system as a base model without any modification to it. To the best of our knowledge, our PriCDSR is the first work to investigate privacy issues in cross-domain sequential recommender systems. We conduct experiments on three domains, and the results demonstrate that our PriCDSR, despite introducing noise, still outperforms recommender systems that only use data from a single domain.

Figures (3)

• Figure 1: Overview of our privacy-preserving cross-domain sequential recommender system (PriCDSR). The plaintext data of the auxiliary domain is represented as a sequential data matrix $\mathbf{R}^A$ (the blue box in the bottom left corner), which consists of the users' historical interaction sequences. $\mathbf{R}^A$ is added with noise to generate the perturbed sequential data matrix $\tilde{\mathbf{R}}^A$ (the orange box in the upper left corner). We incorporate these noises (the orange icons in the orange boxes) using the proposed random mechanism $\mathcal{M}$ (see Fig. \ref{['alg']}) satisfying Sequential DP (see Definition \ref{['def:sdp']}). Subsequently, the perturbed sequential data matrix of the auxiliary domain is transmitted to the target domain (the orange box in the upper right corner). The target domain can use any CDSR method as the base model, and legally use the perturbed sequential data matrix of the auxiliary domain $\tilde{\mathbf{R}}^A$ and the plaintext sequential data matrix of the target domain $\mathbf{R}^T$ (the green box in the bottom right corner) to provide recommendation services.
• Figure 2: Our random algorithm $\mathcal{M}$
• Figure 3: Illustration of an example of our random mechanism $\mathcal{M}$. From top to bottom, each interaction sequence shows each step of how our random mechanism $\mathcal{M}$ adds noise to the example interaction sequence. The item ID sampled at each step is represented by a dotted box (the dotted arrow extending from it points to the current item ID (or padded zero)). If the sampled item ID appears in the subsequent subsequence (connected by a dotted line), the current item ID (or padded zero) and the subsequent item ID will be swapped. Otherwise, the current item ID is replaced with the sampled item ID. Based on the sampled item ID, the current item ID, and whether the sampled item ID appears in the subsequent subsequence, the added noise can be classified into four categories, i.e., insertion, deletion, replacement, and swap. The specific operation of each step is explained on the right side of the corresponding sequence.

Theorems & Definitions (5)

• Definition 1: Sequential Data Matrix
• Definition 2: Neighbouring Sequential Data Matrices
• Definition 3: Sequential Differential Privacy, SDP
• Theorem 1
• proof