Scholarlink

<p>To address the issues of model interpretability and understanding the influence of features, neural additive models were used to decompose predictions into feature-specific additive components. However, they assumed that all input features were equally relevant, and therefore the lack of an explicit feature selection mechanism limited their reliability and interpretability under conditions of high dimensionality or low informativeness of features. We propose a&#xa0;modified neural additive model with integrated hard attention mechanism to select features under processing. Attention weights were learned jointly with feature subnetworks and constrained to produce discrete selection behavior, enabling explicit feature selection in the prediction process. As a&#xa0;result, highly collinear or weakly informative features were suppressed, while informative features retained interpretable functional effects. The proposed mechanism yielded sparse feature utilization patterns, consequently improving clarity of explanations. Quantitative evaluation demonstrated that the proposed model reduced input dimensionality by approximately 40–90% across all datasets while maintaining or improving predictive performance. Thus, this method is well suited for areas, such as clinical risk prediction and diagnostic support, where interpretability of results is important and compact, transparent models are needed. Remaining challenges included sensitivity of sparsity and performance to attention regularization and computational cost at training time, which suggested directions for further research.</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

Differentiable feature selection method for neural additive models

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Rinat Dumaev,
  • Sergei Molodyakov

摘要

To address the issues of model interpretability and understanding the influence of features, neural additive models were used to decompose predictions into feature-specific additive components. However, they assumed that all input features were equally relevant, and therefore the lack of an explicit feature selection mechanism limited their reliability and interpretability under conditions of high dimensionality or low informativeness of features. We propose a modified neural additive model with integrated hard attention mechanism to select features under processing. Attention weights were learned jointly with feature subnetworks and constrained to produce discrete selection behavior, enabling explicit feature selection in the prediction process. As a result, highly collinear or weakly informative features were suppressed, while informative features retained interpretable functional effects. The proposed mechanism yielded sparse feature utilization patterns, consequently improving clarity of explanations. Quantitative evaluation demonstrated that the proposed model reduced input dimensionality by approximately 40–90% across all datasets while maintaining or improving predictive performance. Thus, this method is well suited for areas, such as clinical risk prediction and diagnostic support, where interpretability of results is important and compact, transparent models are needed. Remaining challenges included sensitivity of sparsity and performance to attention regularization and computational cost at training time, which suggested directions for further research.