Bias-Aware Hierarchical Clustering for detecting the discriminated groups of users in recommendation systems

Abstract

One challenge for the modern recommendation systems is the Tyranny of Majority -the generated recommendations are often optimized for the mainstream trends so that the minority preference groups remain discriminated. Moreover, most modern recommendation techniques are characterized as black-box systems. Given a lack of understanding of the dataset characteristics and insufficient diversity of represented individuals, such approaches inevitably lead to amplifying hidden data biases and existing disparities. In this research, we address this problem by proposing a novel approach to detecting and describing potentially discriminated user groups for a given recommendation algorithm. We propose a Bias-Aware Hierarchical Clustering algorithm that identifies user clusters based on latent embeddings constructed by a black-box recommender to identify users whose needs are not met by the given recommendation method. Next, a post-hoc explainer model is applied to reveal the most important descriptive features that characterize these user segments. Our method is model-agnostic and does not require any a priori information about existing disparities and sensitive attributes. An experimental evaluation on a synthetic dataset and two real-world datasets from different domains shows that, compared with other clustering methods and arbitrarily selected user groups, our method is capable of identifying underperforming segments for different recommendation algorithms, and detect more severe disparities.

Fairness in Recommender Systems: Research Landscape and Future Directions

“…• Recommendation quality measures for groups: In a number of works on user-related fairness, the goal is to ensure that no group of users is discriminated by receiving recommendations of lower quality than another (privileged or majority) group. To evaluate this aspect, common accuracy measures can be applied and compared across groups, for example, NDCG [117,35] for user groups, or relevance [68] for item groups. • Recommendation quality measures for individuals: In works that aim at individual fairness-these are mostly works on group recommendation-a common goal often is to make recommendations that are acceptable for all group members, i.e., where the preferences of none of the group members is ignored.…”

“…In terms of fairness towards users, the technical proposal in [75] for example aims to serve users with recommendations that reflect their past diversity preferences with respect to movie genres. An approach towards fairness to groups is proposed in [117]. Here, groups are not identified by their protected attribute, but by the recommendation accuracy that is achieved (using any metric) for the members of the group.…”

“…These could for example be the problem of job recommendations on business networks or people recommendations on social media which can be discriminatory. In today's research, often datasets from rather non-critical domains or synthetic datasets are used to showcase the effectiveness of a technical solution [78,63,118,117,58,43,79,46,119].…”

“…Popularity Bias [77,36] Popularity Count [77] Exposure variance [105] Exposure/visibility gain [57] Exposure/relevance ratio [105] Weighted proportional fairness [73] Fraction of satisfied producers, inequality in exposures distribution [104] Popularity rate (and long-tail rate) [78] Gini index of: recommendation frequency [63]; Popularity scores [79] Ranking-based statistical parity, ranking-based equal-opportunity [127] Divergence or deviation based metrics KL divergence of exposure, normalized discounted KL divergence [72] Bias disparity [43,44,45] Generalized Cross Entropy [35] Max individual deviation, total variation distance, and KL-divergence (between distributions of item interactions) [129] Mean Absolute Difference (MAD) [52,35,36] Group Average Popularity (GAP) [76,130] User/item deviation cost [125] Disparity of exposure [68] Recommendation quality measures for groups NDCG (user groups) [117,59] F1 score [59] F-statistic [41] Relevance (item groups) [68] MAD-ranking [35] Recommendation quality measures for individuals…”

Approaches for the Clustering of Geographic Metadata and the Automatic Detection of Quasi-Spatial Dataset Series

“…To identify similarities by period, theme, or researcher, they cluster the works in the analysed collection using Euclidean distance similarity, Partitioning Around Medoids, and K-means. Misztal-Radecka and Indurkhya [48] describes a Bias-Aware Hierarchical Clustering algorithm to improve recommendation systems by identifying clusters of users with unsuitable recommendations. It is a variation of K-means where splitting depends on high biases instead of minimum variance.…”