Scholarlink

<p>The rise of AI in marketing engenders crucial questions of equity in customer segmentation wherein biased data may exclude groups that are vulnerable to discrimination. Can generative AI solve this problem? We attempt to solve these issues with a Bias-Aware Generative AI (BiA-GAI) framework, using the Mall Customer Segmentation dataset. After normalizing the dataset with z-score and min-max normalization, we proceed to train a Variational Autoencoder (VAE) with an encoder (64-32-2) and decoder (2-32-64) architecture that includes dropout (0.2) alongside adversarial training to reduce bias arising from the protected attribute (income ≥ 50). Clustering of the debiased latent representations with BIRCH (K = 4) yields Davies-Bouldin Index (DBI): 0.85 (z-score) and 0.92 (min-max), lower than those of K-Means (DBI: 1.12). From the standpoint of fairness, near-perfect parity is achieved. Disparate Impact reaches 0.999 (plus 22.4% from baseline 0.816), while Mean Difference remains at almost zero (-1.11e-16 for z-score, -2.45e-15 for min-max). Tested on the UCI Adult dataset, BiA-GAI scores a Disparate Impact of 0.995, asserting its robustness. This approach offers actionable insights for ethical marketing strategies, addressing a pressing need in AI-driven consumer analytics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bias-Aware Generative AI for fair customer segmentation: a VAE and BIRCH clustering approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月15日

查看全文
  • Shivranjani Bharatbhai Gajjar,
  • Vivekanand Mishra

摘要

The rise of AI in marketing engenders crucial questions of equity in customer segmentation wherein biased data may exclude groups that are vulnerable to discrimination. Can generative AI solve this problem? We attempt to solve these issues with a Bias-Aware Generative AI (BiA-GAI) framework, using the Mall Customer Segmentation dataset. After normalizing the dataset with z-score and min-max normalization, we proceed to train a Variational Autoencoder (VAE) with an encoder (64-32-2) and decoder (2-32-64) architecture that includes dropout (0.2) alongside adversarial training to reduce bias arising from the protected attribute (income ≥ 50). Clustering of the debiased latent representations with BIRCH (K = 4) yields Davies-Bouldin Index (DBI): 0.85 (z-score) and 0.92 (min-max), lower than those of K-Means (DBI: 1.12). From the standpoint of fairness, near-perfect parity is achieved. Disparate Impact reaches 0.999 (plus 22.4% from baseline 0.816), while Mean Difference remains at almost zero (-1.11e-16 for z-score, -2.45e-15 for min-max). Tested on the UCI Adult dataset, BiA-GAI scores a Disparate Impact of 0.995, asserting its robustness. This approach offers actionable insights for ethical marketing strategies, addressing a pressing need in AI-driven consumer analytics.