Scholarlink

With its profound potential, causal inference stands as a powerful tool offering invaluable insights that can inform and enhance the development of effective strategies. However, existing approaches often fall short by neglecting latent confounding factors or oversimplifying the complex interactions between observable behaviors and underlying socioeconomic mechanisms. To address these limitations, we propose a robust causal inference framework, the Heterogeneous Causal Effects Variational Autoencoder (HCEVAE). The HCEVAE framework comprises two core components: a factor perception encoder and a counterfactual predictor. The factor perception encoder jointly models observed treatments and unmeasured mediating factors, effectively bridging explicit observables and latent confounders to capture complex dependencies. Meanwhile, the counterfactual predictor, equipped with mixed-control mechanisms, integrates observed variables and latent confounders to simulate outcomes under alternative scenarios. Additionally, we design a heterogeneity estimator to dynamically classify features into indirect confounding variables and direct causal variables, enabling precise confounding control and personalized causal effect analysis. Experimental validation on the popular public dataset IHDP demonstrates that, compared to the best baseline model, HCEVAE reduces heterogeneous effect estimation error (PEHE) by 12.6%. Furthermore, a real-world case study using China General Social Survey (CGSS) data (2015–2021) also reveals the framework’s capability to disentangle complex causal pathways.

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

HCEVAE: A Robust Heterogeneous Causal Effects Variational Autoencoder Framework

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Mengjia Yi,
  • Haiyong Shi,
  • Keqin Zhong,
  • Guangwei Xia,
  • Bingyi Liu

摘要

With its profound potential, causal inference stands as a powerful tool offering invaluable insights that can inform and enhance the development of effective strategies. However, existing approaches often fall short by neglecting latent confounding factors or oversimplifying the complex interactions between observable behaviors and underlying socioeconomic mechanisms. To address these limitations, we propose a robust causal inference framework, the Heterogeneous Causal Effects Variational Autoencoder (HCEVAE). The HCEVAE framework comprises two core components: a factor perception encoder and a counterfactual predictor. The factor perception encoder jointly models observed treatments and unmeasured mediating factors, effectively bridging explicit observables and latent confounders to capture complex dependencies. Meanwhile, the counterfactual predictor, equipped with mixed-control mechanisms, integrates observed variables and latent confounders to simulate outcomes under alternative scenarios. Additionally, we design a heterogeneity estimator to dynamically classify features into indirect confounding variables and direct causal variables, enabling precise confounding control and personalized causal effect analysis. Experimental validation on the popular public dataset IHDP demonstrates that, compared to the best baseline model, HCEVAE reduces heterogeneous effect estimation error (PEHE) by 12.6%. Furthermore, a real-world case study using China General Social Survey (CGSS) data (2015–2021) also reveals the framework’s capability to disentangle complex causal pathways.