Fairness in Recommendation Systems: Consumer and Provider Perspectives

The presentation introduces the theme of fairness in recommendation systems, outlining a two‑sided analysis: the consumer (user) perspective and the provider (supplier) perspective.

It first notes that recommendation systems are ubiquitous in entertainment, news, shopping, etc., and highlights concerns that these systems can both reflect and shape user preferences, potentially introducing bias that harms trust and sustainability.

Three primary sources of bias are identified: limited user attention, limited recommendation slots, and biased training data, which can lead to phenomena such as the Matthew effect, filter bubbles, and market imbalance.

Fairness is defined in two common ways: Equality (equal opportunity for all users) and Equity (adjusted support for disadvantaged groups). An example contrasts gender equality in hiring with equity for disabled applicants.

From the consumer side, the goal is to ensure equal treatment across sensitive groups, using metrics like Counterfactual Group Fairness (CGF) to measure disparity. Experiments on the MovieLens dataset illustrate how different distributions of a sensitive attribute (e.g., position) affect fairness outcomes.

From the provider side, fairness means supporting smaller suppliers to avoid monopolies, measured by Max‑min Fairness (MMF) which aims to improve the utility of the worst‑off providers. Analyses on the Yelp dataset confirm the presence of such imbalances.

The article proposes a causal embedding model that separates user embeddings into sensitive‑related and insensitive components using instrumental variable regression. The CGF metric is incorporated as a regularizer, and the resulting loss function guides training toward fairer recommendations.

Optimization is performed in the dual space to enable efficient online solving, with both CPU and GPU implementations showing stable inference time as candidate sets grow. The solution’s theoretical bounds are also discussed.

In conclusion, a multi‑role recommendation system must guarantee equality for users and fairness for providers; CGF and MMF serve as respective measurement tools, and model adjustments can improve both aspects. Future work includes cross‑platform fairness, integrating market‑level considerations, and balancing exposure with quality control.