Transparency, Privacy, and Fairness in Recommender Systems

TL;DR

This work investigates how transparency, privacy, and fairness can be embedded into recommender systems by combining psychology-informed cognitive models with modern privacy techniques and fairness analyses. It presents seven scientific contributions across three themes: cognitive-model-based transparency (MINERVA2, ACT-R), privacy under limited preference information (differential privacy, ReuseKNN, session-based/cold-start methods), and popularity-bias–driven fairness (measuring, mitigating, and understanding long-term dynamics). The research demonstrates concrete methods (e.g., SoMe tagging, ReuseKNN, variational autoencoders for session data) and provides reproducible resources (code, datasets, DOIs) to advance trustworthy RS. The findings highlight practical implications for policy and system design, including improved transparency through cognitive explanations, stronger privacy guarantees with targeted DP protection, and nuanced fairness insights across user groups and domains, such as news and labor-market settings.

Abstract

Recommender systems have become a pervasive part of our daily online experience, and are one of the most widely used applications of artificial intelligence and machine learning. Therefore, regulations and requirements for trustworthy artificial intelligence, for example, the European AI Act, which includes notions such as transparency, privacy, and fairness are also highly relevant for the design of recommender systems in practice. This habilitation elaborates on aspects related to these three notions in the light of recommender systems, namely: (i) transparency and cognitive models, (ii) privacy and limited preference information, and (iii) fairness and popularity bias in recommender systems. Specifically, with respect to aspect (i), we highlight the usefulness of incorporating psychological theories for a transparent design process of recommender systems. We term this type of systems psychology-informed recommender systems. In aspect (ii), we study and address the trade-off between accuracy and privacy in differentially-private recommendations. We design a novel recommendation approach for collaborative filtering based on an efficient neighborhood reuse concept, which reduces the number of users that need to be protected with differential privacy. Furthermore, we address the related issue of limited availability of user preference information, e.g., click data, in the settings of session-based and cold-start recommendations. With respect to aspect (iii), we analyze popularity bias in recommender systems. We find that the recommendation frequency of an item is positively correlated with this item's popularity. This also leads to the unfair treatment of users with little interest in popular content. Finally, we study long-term fairness dynamics in algorithmic decision support in the labor market using agent-based modeling techniques.

Figures (8)

• Figure 1: An example illustrating the difference between the BLL equation (left panel) and the activation equation (right panel). Here, unfilled nodes represent target genres $g_1$ and $g_2$, and black nodes represent contextual genres. For $g_1$ and $g_2$, the node sizes represent the activation levels, and for the contextual genres, the node sizes represent the weights $W_c$. The association strength $S_{c,g}$ is represented by each edge's length. We see a different ranking of the genres in the two settings, which illustrates the importance of the associative activation kowald2015refiningtrattner2016modelingkowald2020utilizing.
• Figure 2: Schematic illustration of the data usage (i.e., how often a user is used as a neighbor) distribution of UserKNN and ReuseKNN. ReuseKNN increases the number of secure users (green, no differential privacy needed) and decreases the number of vulnerable users (red, differential privacy needs to be applied) compared to UserKNN. The dashed line illustrates the data usage threshold $\tau$, a parameter to adjust the maximum data usage for users to be treated as secure.
• Figure 3: (a) Calculation of the BLL equation's $d$ parameter. On a log-log scale, we plot the relistening count of the genres over the time since their last listening event (LE), and set $d$ to the slopes $\alpha$ of the linear regression lines lex2020modeling. (b) Recall/precision plots for $k = 1 \ldots 10$ predicted genres of the baselines, and our $BLL_u$ and $ACT_{u,a}$ approaches. $ACT_{u,a}$ achieves the highest accuracy kowald2020utilizing.
• Figure 4: Heatmap illustrating the relative contribution of three ACT-R components (BLL, S, and V) and one social component (SC) to the recommendation scores of six recommended tracks for a randomly chosen Last.fm user recsys_actr_2023.
• Figure 5: Mean absolute error (MAE) of neural-based KNN recommender system variants. Our results indicate that combining neighborhood reuse with differential privacy (NeuKNN+Reuse$_{DP}$) yields better accuracy (lower MAE) than neural-based methods that do not apply neighborhood reuse (NeuKNN$^{full}_{DP}$) tist_dp_2023.