A Review of Causal Fairness in E-commerce Recommendation Systems: From Behavior Disentanglement to Counterfactual Intervention

Abstract

Within online retail, e-commerce recommendation systems occupy a study-defining role, yet they have drawn growing criticism for producing stereotyped or biased outcomes. Conventional fairness metrics —notably demographic parity and equalized odds —treat ever y observed correlation between user attributes and recommendation results as comparably problematic; what such metrics do not do, however, is separate legitimate personalization from unfair discrimination. Taking up that still-unresolved distinction, this review interrogates how far theories of causal inference can, in practice, discriminate between the two cases. Employing a narrative literature review design, the paper synthesizes scholarship across three complementary causal frameworks: the Structural Ca usal Model (SCM), Path -Specific Effects (PSE), and Causal Mediation Analysis (CMA). At the level of the consolidated findings, SCM is shown to separate user behavior into intrinsic preference, item visibility, and conformity effects, with only the first co nstituting a legitimate basis for personalization. When the same problem is recast through PSE decomposition, unfair bias is located in the visibility and conformity pathways rather than in preference pathways. From CMA, meanwhile, the estimates indicate that contextual variables, including browsing history and social popularity, account for 15-30% of the total bias effect. Set alongside one another, these findings yield the first systematic synthesis of the three frameworks as applied specifically to fairn ess evaluation in e-commerce recommendation systems, thereby furnishing a technical blueprint for platform-level causal fairness audits.