Abstract
A growing proportion of human interactions are digitized on social media platforms and subjected to algorithmic decision-making, and it has become increasingly important to ensure fair treatment from these algorithms. In this work, we investigate gender bias in collaborative-filtering recommender systems trained on social media data. We develop neural fair collaborative filtering (NFCF), a practical framework for mitigating gender bias in recommending career-related sensitive items (e.g. jobs, academic concentrations, or courses of study) using a pre-training and fine-tuning approach to neural collaborative filtering, augmented with bias correction techniques. We show the utility of our methods for gender de-biased career and college major recommendations on the MovieLens dataset and a Facebook dataset, respectively, and achieve better performance and fairer behavior than several state-of-the-art models.
Introduction
Recommender systems are ubiquitous, but when they are used for "consequential" life decisions—like suggesting a job or a college major—their tendency to mirror and amplify societal biases becomes a serious ethical and legal risk. For example, if a model sees that 97% of "homemakers" in its dataset are women, it may stop recommending that path to men, or conversely, fail to recommend STEM careers to women due to historical underrepresentation.
This research addresses two main hurdles in fair career recommendation: data sparsity (users typically have only one career or major) and implicit bias (models can infer protected attributes like gender from "non-sensitive" proxy data, such as movie likes).
Methodology
The proposed Neural Fair Collaborative Filtering (NFCF) framework utilizes a three-step process to ensure accurate yet equitable recommendations:
- Pre-training: A Deep Neural Network (DNN) is first trained on massive amounts of non-sensitive implicit feedback data (e.g., movie ratings or page likes). This allows the model to learn robust user latent vectors (embeddings) despite the scarcity of career-specific data.
- Embedding De-biasing: To prevent the model from using proxies for gender, a linear projection technique is applied. A "gender bias vector" is calculated from the average embeddings of male and female users and then subtracted from every user's latent vector.
- Fine-tuning with a Fairness Penalty: The model is fine-tuned on sensitive career items. During this phase, a Differential Fairness (DF) penalty is added to the loss function, regularizing the model to ensure recommendation probabilities do not differ excessively between genders.
Experimental Results
The framework was evaluated on the MovieLens (1M) dataset (using movie ratings to predict 17 occupations) and a Facebook dataset (using page likes to predict 169 college majors).
Performance and Equity
NFCF consistently outperformed state-of-the-art fair baselines in both accuracy and equity. On the MovieLens dataset, NFCF achieved a Hit Ratio (HR@5) of 0.670, nearly identical to the biased version, while cutting the mean differential fairness measure ($\epsilon_{mean}$) by more than half (from 0.188 to 0.083). On the Facebook dataset, the model significantly reduced absolute unfairness while maintaining high ranking quality as measured by Normalized Discounted Cumulative Gain (NDCG).
Qualitative Impact
In the Facebook dataset, the percentage of women recommended for "computer science" rose from 27% in the raw data to 48% with NFCF, moving the results significantly closer to demographic parity.
Discussion and Key Insights
An ablation study revealed a "double-correction" necessity: neither embedding debiasing nor the fairness penalty is sufficient on its own. Addressing bias at both the input (embeddings) and output (prediction) levels is required for the best balance of fairness and accuracy. Interestingly, the fairness penalty sometimes improved accuracy, suggesting that societal bias in data can act as "noise" that, when removed, allows the model to generalize better to true underlying user preferences.
While this work focuses on binary gender, future research will explore intersectional fairness (e.g., race and age) and adversarial networks to automatically decouple user embeddings from multiple protected attributes simultaneously.
Summary updated