Abstract:Fairness-aware recommendation alleviates discrimination issues to build trustworthy recommendation systems. Explaining the causes of unfair recommendations is critical, as it promotes fairness diagnostics, and thus secures users’ trust in recommendation models. Existing fairness explanation methods suffer high computation burdens due to the large-scale search space and the greedy nature of the explanation search process. Besides, they perform feature-level optimizations with continuous values, which are not applicable to discrete attributes such as gender and age. In this work, we adopt counterfactual explanations from causal inference and propose to generate attribute-level counterfactual explanations, adapting to discrete attributes in recommendation models. We use real-world attributes from Heterogeneous Information Networks (HINs) to empower counterfactual reasoning on discrete attributes. We propose a Counterfactual Explanation for Fairness (CFairER) that generates attribute-level counterfactual explanations from HINs for item exposure fairness. Our CFairER conducts off-policy reinforcement learning to seek high-quality counterfactual explanations, with attentive action pruning reducing the search space of candidate counterfactuals. The counterfactual explanations help to provide rational and proximate explanations for model fairness, while the attentive action pruning narrows the search space of attributes. Extensive experiments demonstrate our proposed model can generate faithful explanations while maintaining favorable recommendation performance.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is the fairness issue in recommendation systems, especially explaining the causes leading to unfair recommendations. Specifically, the author focuses on how to generate counterfactual explanations at the attribute level to help understand why certain recommendation results are unfair and provide methods to improve these unfairness.
### Problem Background
Recommendation systems (RS) are widely used in many online services, such as e - commerce, social media, etc. It plays a role in resource allocation by distributing content (such as products, articles, etc.) to multi - stakeholder (users and content providers). However, whether the allocation of recommendation systems is fair, that is, recommendation fairness, has become an important problem that needs to be solved urgently. Research shows that recommendation systems may suffer from various unfair problems, such as favoring popular items while ignoring niche items, leading to the Matthew effect and other problems.
### Limitations of Existing Methods
Existing fairness explanation methods are mainly divided into factual explainable methods and counterfactual explainable methods. Factual explanation methods explain the unfairness of the model by estimating the importance of features to the fairness of the model, but this method has the following limitations:
1. **High computational burden**: Since all possible feature combinations need to be considered, the search space grows exponentially.
2. **Only consider relative importance**: Removing these features does not necessarily reduce the unfairness of the model.
Counterfactual explanation methods ensure the reduction of unfairness by learning the "minimal" changes of input features and provide "if - then" explanations. However, existing methods are mainly optimized for continuous features and cannot handle discrete attributes (such as gender, age, etc.).
### Solutions in the Paper
To solve the above problems, this paper proposes a counterfactual explanation method based on heterogeneous information networks (HINs) (Counterfactual Explanation for Fairness, CFairER), aiming to generate counterfactual explanations at the attribute level. The specific contributions are as follows:
1. **Utilize rich attributes in HINs**: Conduct counterfactual reasoning through real - world attributes in HINs (such as the gender of users, the category of items, etc.), overcoming the limitation that existing methods can only handle continuous features.
2. **Introduce an off - policy reinforcement learning framework**: Use reinforcement learning to identify the optimal counterfactual explanation and prune the search space through the attention mechanism to reduce the computational burden.
3. **Design a counterfactual risk minimization objective**: Ensure that the generated explanations are credible and faithful to the change in model fairness.
### Method Overview
The CFairER framework contains three main components:
1. **Graph representation module**: Embed users, items, and attributes into dense vectors for state representation learning and attention - action pruning.
2. **Recommendation model**: Learn the latent factors of users and items and generate recommendation results.
3. **Counterfactual fairness explanation model (CFE)**: Combine the graph representation module and the recommendation model for counterfactual reasoning and generate counterfactual explanations at the attribute level.
In this way, CFairER can not only explain the unfairness of the recommendation system but also provide specific methods to improve these unfairness, thereby increasing the credibility and user trust of the recommendation system.
### Conclusion
This paper solves the limitations of existing methods in handling discrete attributes by introducing the counterfactual explanation method at the attribute level, and effectively reduces the computational burden through reinforcement learning and the attention mechanism, improving the reliability and effectiveness of the explanation.