Addressing Consumers' Sensitive Attributes in Product Recommendations: An Explainable AI Recommendations System Approach

Sensitive consumer attributes such as race are an important concern for marketing managers for their recommendation systems. Leveraging race to make recommendations for technology products runs counter to recent regulatory ask of reducing reliance on such attributes, e.g., for internet and broadband offerings (FCC, 2023). This work proposes an explainable AI approach to product recommendations that can handle sensitive attributes. Specifically, it advances the Xian et al. (2019) knowledge graph (KG) and policy-guided path reasoning approach with a regularization framework to address sensitive attributes. We demonstrate the efficacy of the proposed approach on products in the cellphones category using customer reviews data from Amazon. The proposed approach provides near-highest recommendation quality (compared to existing benchmarks) and outperforms on decorrelating sensitive attributes with recommendations. The proposed approach is also flexible in handling sensitive attributes. We examine the Beauty category as an extension, where it leverages race and outperforms benchmarks on recommendation quality. Explaining reasons for product recommendations is straightforward with the proposed approach. Marketing managers will find the proposed approach useful as regulators and consumers increasingly demand explainability for firms' recommendation decisions.