The Impact of Differential Privacy on Recommendation Accuracy and Popularity Bias

TL;DR

This paper investigates how applying Differential Privacy (DP) to training data affects personalized recommendations in collaborative filtering. It implements the DP mechanism for implicit feedback and evaluates three state-of-the-art models across three datasets over a range of privacy budgets $\epsilon$, demonstrating that DP changes nearly all users' recommendations, reduces recommendation accuracy, and increases popularity bias, especially at lower $\epsilon$. A key finding is the 'poor get poorer' effect, where users who prefer unpopular items experience the strongest DP-induced bias, highlighting a crucial privacy-accuracy-popularity trade-off. The results inform privacy-preserving recommender design and motivate group-aware mitigation strategies to balance user privacy with personalization quality and item popularity fairness.

Abstract

Collaborative filtering-based recommender systems leverage vast amounts of behavioral user data, which poses severe privacy risks. Thus, often, random noise is added to the data to ensure Differential Privacy (DP). However, to date, it is not well understood, in which ways this impacts personalized recommendations. In this work, we study how DP impacts recommendation accuracy and popularity bias, when applied to the training data of state-of-the-art recommendation models. Our findings are three-fold: First, we find that nearly all users' recommendations change when DP is applied. Second, recommendation accuracy drops substantially while recommended item popularity experiences a sharp increase, suggesting that popularity bias worsens. Third, we find that DP exacerbates popularity bias more severely for users who prefer unpopular items than for users that prefer popular items.

Figures (2)

• Figure 1: DP's impact on recommendation accuracy as measured by $\Delta$Recall. DP leads to a severe drop in recommendation accuracy. In particular, this drop becomes more serious for small $\epsilon$ values that provide a high level of privacy. This corresponds to the well-known accuracy-privacy trade-off (RQ2).
• Figure 2: DP's impact on popularity bias as measured by $\Delta ARP$ and $\Delta PopLift$. We find that DP increases ARP, which becomes more severe the smaller the $\epsilon$ value is (see Figure \ref{['fig:delta_arp']}). Plus, the recommendation popularity mismatches the item popularity distribution in the user profiles (see Figure \ref{['fig:delta_pl']}). Overall, this gives strong evidence that DP makes popularity bias worse (RQ3).