Scholarlink

<p>Estimating causal effects from real-world observational data, which often mixes structured features and unstructured text, is crucial for data-driven decision-making. However, existing methodologies face a fundamental triple challenge: balancing performance with interpretability, and establishing credible validation without randomized controlled trials. We propose <b>MixingInsights</b>, a framework that introduces a dual-path architecture for learning confounder representations, along with a systematic validation protocol. One path constructs interpretable proxy variables by employing a keyword-assisted topic model to extract semantically coherent concepts from text, which are then enriched with sentiment dimensions to form a transparent representation. The second path focuses on learning a balanced, multimodal representation by jointly encoding text and structured features through a neural network, aiming to approximate the conditions of ignorability. Validation employs semi-synthetic benchmarks, comparison to standard text models, and alignment with domain expertise. The framework accurately recovers causal effects on semi-synthetic benchmarks, outperforming models using only structured data or standard text representations. In real airline reviews, it confirms known factors (e.g., seat comfort) and reveals a new driver of satisfaction, namely, perceived price value. <b>MixingInsights</b> tackles key challenges in causal inference with mixed data by combining interpretable and deep representation paths with rigorous validation. It supports both the confirmation of domain knowledge and the discovery of novel, actionable insights, advancing practical data-driven decision-making.</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

MixingInsights: A Framework for Causal Inference with Confounder Representation Learning from Mixed Structured and Textual Data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • Lu Feng,
  • Adi Lin,
  • Zhiyong Luo

摘要

Estimating causal effects from real-world observational data, which often mixes structured features and unstructured text, is crucial for data-driven decision-making. However, existing methodologies face a fundamental triple challenge: balancing performance with interpretability, and establishing credible validation without randomized controlled trials. We propose MixingInsights, a framework that introduces a dual-path architecture for learning confounder representations, along with a systematic validation protocol. One path constructs interpretable proxy variables by employing a keyword-assisted topic model to extract semantically coherent concepts from text, which are then enriched with sentiment dimensions to form a transparent representation. The second path focuses on learning a balanced, multimodal representation by jointly encoding text and structured features through a neural network, aiming to approximate the conditions of ignorability. Validation employs semi-synthetic benchmarks, comparison to standard text models, and alignment with domain expertise. The framework accurately recovers causal effects on semi-synthetic benchmarks, outperforming models using only structured data or standard text representations. In real airline reviews, it confirms known factors (e.g., seat comfort) and reveals a new driver of satisfaction, namely, perceived price value. MixingInsights tackles key challenges in causal inference with mixed data by combining interpretable and deep representation paths with rigorous validation. It supports both the confirmation of domain knowledge and the discovery of novel, actionable insights, advancing practical data-driven decision-making.