Cost-sensitive meta-learning can effectively balance exposure fairness in recommendation systems without compromising their utility.

In our paper, which will be presented at the ECIR 2024 conference, we introduce a novel cost-sensitive meta-learning technique aimed at enhancing fairness in recommendation systems. Our work addresses a critical issue in many online platforms – ensuring equitable exposure for all provider groups, including those typically underrepresented. This work is a collaborative effort with Giulia Cerniglia, Mirko Marras, Alessandra Perniciano, and Barbara Pes.

Recommender systems are ubiquitous, guiding users in e-commerce, online streaming, and even educational platforms. Traditionally, these systems focus on optimizing consumer experiences, often overlooking the needs and fair representation of content providers. This oversight can lead to an imbalance in exposure, particularly disadvantaging newcomers and minority groups. Recognizing this, we aimed to create a solution that ensures a fair distribution of exposure, aligning with the principle of equity.

Our Solution: A Cost-Sensitive Meta-Learning Technique

Our approach is a cost-sensitive meta-learning technique that effectively manages the exposure of different provider groups in recommendation lists. By manipulating the sampling distribution in the training data, we can control and balance the exposure that various groups receive. This method is not only effective but also adaptable, thanks to its meta-learning nature, making it compatible with various underlying recommendation algorithms.

How It Works

1. Fairness Objective Definition. We started by defining our fairness objective – to secure an equitable distribution of exposure for provider groups relative to their representation within the catalog. This means if a group makes up a certain percentage of the catalog, they should ideally receive a proportional percentage of exposure in recommendation lists.
2. Data Preparation and Model Creation. Using publicly available datasets, we prepared our data, ensuring a fair representation of provider groups. We then employed a pairwise learning model, specifically the Bayesian-Personalized Ranking with Matrix Factorization (BPRMF), known for its effectiveness in personalized ranking algorithms.
3. Traditional vs. Cost-Sensitive Sampling. We first established a baseline using traditional pairwise training via random sampling. Then, we introduced our cost-sensitive sampling approach, adjusting the distribution of items from different provider groups in the training triplets. This adjustment allowed us to control the balance between groups, aiming to reflect the target fairness objective in the resulting recommendations.

Our Findings

Our evaluation showed promising results. By adjusting the parameter in our cost-sensitive approach, we successfully balanced the representation of minority groups in both positive and negative item sets, achieving an exposure rate that closely matches their representation in the catalog. We achieved this without sacrificing the utility of the recommendation system – a common challenge in incorporating fairness objectives.

Conclusion

In conclusion, our research presents a significant step forward in making recommendation systems fairer and more equitable. By introducing a novel cost-sensitive meta-learning approach, we’ve shown that it’s possible to balance fairness and utility, ensuring that all provider groups receive the exposure they deserve.

Our future work will involve applying our method to other pairwise recommendation methods and conducting extensive performance evaluations across diverse domains. We aim to demonstrate the adaptability and effectiveness of our approach further, promoting the integration of the cost-sensitive paradigm into recommender systems more broadly.