Gender-typed Preference Bias in Recommender Systems

🔬 Ongoing Research Co-supervised

This research investigates a novel and unexplored question in recommender systems: Does the presence of gender-typed preferences distort or suppress the representation of more neutral preferences within a recommender model?

This is particularly relevant in multi-domain data where strongly gender-targeted items (e.g., cosmetics) exist alongside more general-audience items (e.g., science books). The research examines whether recommender systems inadvertently narrow exposure diversity by over-emphasising demographic signals.

Research Focus

Core Research Question

When recommender models learn from data containing both gender-typed and gender-neutral items, do the gender-typed preferences act as proxies that bias recommendations away from neutral items?

Research Approach

The research follows a systematic methodology:

Identify gender-typed preferences: Examine whether specific items or groups are statistically associated with particular genders
Detect proxy usage: Show that latent gender signals in data are used as proxies within recommendation models
Prove bias exists: Demonstrate that recommendations are biased towards gender-typed items
Analyse impact: Examine whether gender-neutral preferences are suppressed, narrowing exposure diversity
Counterfactual testing: Test whether recommendations for non-gender-typed items improve when trained in isolation

Key Contributions

Impact Identification

Identifying how gender-typed preferences affect neutral item recommendations, revealing potential harm to recommendation diversity.

Reproducible Framework

A reproducible framework for analysing demographic-typed preference biases that can be generalised to other demographics and datasets.

Mitigation Strategies

Strategies for mitigating bias to promote more balanced exposure across user groups and content types.

Broader Implications

While this research focuses on gender as a demographic dimension, the framework and findings have broader implications:

Generalisability: The methodology can be applied to other demographic dimensions where item targeting occurs
Fairness in AI: Contributes to understanding how algorithmic systems can inadvertently reinforce demographic biases
Exposure Diversity: Addresses concerns about filter bubbles and narrowing of user exposure to diverse content
Responsible AI: Provides tools for auditing and improving fairness in recommendation systems

Context

This research sits at the intersection of recommender systems, fairness in machine learning, and computational social science. As recommender systems become increasingly influential in shaping user choices and experiences, understanding and mitigating their biases becomes critical for ensuring equitable outcomes.