Scholarlink

<p>Understanding travel mode choice is essential for effective planning and policy-making, as it strongly influences travel demand. Traditionally, mode choice has been modeled using econometric approaches based on random utility theory, particularly discrete choice models (DCMs) like the widely used Multinomial Logit (MNL) model. However, recent advances in machine learning (ML) have shown that ML models can outperform traditional DCMs in predictive accuracy, though they often lack the interpretability of statistical models. In response, hybrid models combining DCMs with ML techniques have emerged. For instance, Sifringer et al. (Transp. Res. Part B Methodol. 140:236–261) introduced the Learning Multinomial Logit (L-MNL) model, which separates the systematic component of an MNL model into an interpretable part that retains the MNL structure and a non-interpretable (learning) part for attributes that do not require interpretation. This approach, though effective, still lacks full transparency. To address this, our study integrates Shapley Additive Explanations (SHAP) with the DeepLIFT algorithm to enhance interpretability in the L-MNL framework. Using the ”Swissmetro” dataset, our results show that the L-MNL model outperforms the MNL in predictive accuracy, while SHAP elucidates the impact of individual attribute values on choice probabilities. This study concludes that combining L-MNL with SHAP offers a promising path toward interpretable ML models for travel mode choice prediction.</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

Deep SHAP explanations for predicted choice probabilities from learning multinomial logit model in modal choice application

  • 期刊论文

  • 发表日期：

    2025年7月21日

查看全文
  • Exequiel Salazar,
  • Patricio Salas,
  • Patricio Sáez

摘要

Understanding travel mode choice is essential for effective planning and policy-making, as it strongly influences travel demand. Traditionally, mode choice has been modeled using econometric approaches based on random utility theory, particularly discrete choice models (DCMs) like the widely used Multinomial Logit (MNL) model. However, recent advances in machine learning (ML) have shown that ML models can outperform traditional DCMs in predictive accuracy, though they often lack the interpretability of statistical models. In response, hybrid models combining DCMs with ML techniques have emerged. For instance, Sifringer et al. (Transp. Res. Part B Methodol. 140:236–261) introduced the Learning Multinomial Logit (L-MNL) model, which separates the systematic component of an MNL model into an interpretable part that retains the MNL structure and a non-interpretable (learning) part for attributes that do not require interpretation. This approach, though effective, still lacks full transparency. To address this, our study integrates Shapley Additive Explanations (SHAP) with the DeepLIFT algorithm to enhance interpretability in the L-MNL framework. Using the ”Swissmetro” dataset, our results show that the L-MNL model outperforms the MNL in predictive accuracy, while SHAP elucidates the impact of individual attribute values on choice probabilities. This study concludes that combining L-MNL with SHAP offers a promising path toward interpretable ML models for travel mode choice prediction.