Scholarlink

<p>Causal inference serves as a critical framework in diverse domains, including healthcare, economics, and the social sciences. Within this context, data-driven machine learning models have gained prominence for estimating individual treatment effects (ITEs). Recently, the Causal Effect Variational Autoencoder (CEVAE) was proposed to infer ITEs from observational data with unobserved confounders. However, CEVAE faces limitations in certain scenarios for instance, it struggles to accurately infer latent confounders in high-dimensional data, leading to biased causal effect estimates. To address these challenges, we introduce the Transformer Causal Effect Variational Autoencoder (TCE-VAE), a novel architecture that integrates transformer-based encoder-decoders with variational autoencoders. The core innovation of TCE-VAE lies in its use of a self-attention mechanism to directly estimate causal effects by capturing complex dependencies and interactions within the data. This approach enhances robustness in learning latent confounders, particularly in high-dimensional settings. We evaluate TCE-VAE on the Infant Health and Development Program (IHDP) dataset. Experimental results demonstrate that the proposed approach outperforms popular state-of-the-art models in terms of average treatment effect estimation. Furthermore, our method exhibits superior precision and robustness in individual treatment effect estimation compared to existing approaches.</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

Transformer-variational autoencoder for estimating individual treatment effect using causal inference framework

  • 期刊论文

  • 发表日期：

    2025年7月9日

查看全文
  • Sohail Ahmad,
  • Hong Wang

摘要

Causal inference serves as a critical framework in diverse domains, including healthcare, economics, and the social sciences. Within this context, data-driven machine learning models have gained prominence for estimating individual treatment effects (ITEs). Recently, the Causal Effect Variational Autoencoder (CEVAE) was proposed to infer ITEs from observational data with unobserved confounders. However, CEVAE faces limitations in certain scenarios for instance, it struggles to accurately infer latent confounders in high-dimensional data, leading to biased causal effect estimates. To address these challenges, we introduce the Transformer Causal Effect Variational Autoencoder (TCE-VAE), a novel architecture that integrates transformer-based encoder-decoders with variational autoencoders. The core innovation of TCE-VAE lies in its use of a self-attention mechanism to directly estimate causal effects by capturing complex dependencies and interactions within the data. This approach enhances robustness in learning latent confounders, particularly in high-dimensional settings. We evaluate TCE-VAE on the Infant Health and Development Program (IHDP) dataset. Experimental results demonstrate that the proposed approach outperforms popular state-of-the-art models in terms of average treatment effect estimation. Furthermore, our method exhibits superior precision and robustness in individual treatment effect estimation compared to existing approaches.