Scholarlink

With the advancement of Artificial Intelligence (AI) techniques in different areas of our society, namely with the use of Machine and Deep Learning models, some challenges must be faced. One of these challenges is responding to the lack of transparency in these models, which makes it difficult to explain the results they obtained. This research centres on predicting the concentration of nitrogen dioxide (NO \({_2}\) ), a critical air pollutant, using a Long Short-Term Memory model (LSTM) along with applying Explainable AI (XAI) techniques to understand the predictions made. Two explainability techniques were applied: an ante-hoc approach with an Attention layer and a post-hoc approach using Shapley Additive Explanations (SHAP). The Attention layer identified carbon monoxide (CO) and NO \({_2}\) as the most powerful features, while the SHAP analysis highlighted NO \({_2}\) as the predominant contributor to the predictions, followed by particulate matter (PM \(_{2.5}\) ) and CO. The results demonstrated that the target value significantly impacts the model’s forecast for both XAI techniques.

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

Employing Explainable AI Techniques for Air Pollution: An Ante-Hoc and Post-Hoc Approach in Dioxide Nitrogen Forecasting

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Pedro Oliveira,
  • Francisco Franco,
  • Afonso Bessa,
  • Dalila Durães,
  • Paulo Novais

摘要

With the advancement of Artificial Intelligence (AI) techniques in different areas of our society, namely with the use of Machine and Deep Learning models, some challenges must be faced. One of these challenges is responding to the lack of transparency in these models, which makes it difficult to explain the results they obtained. This research centres on predicting the concentration of nitrogen dioxide (NO \({_2}\) ), a critical air pollutant, using a Long Short-Term Memory model (LSTM) along with applying Explainable AI (XAI) techniques to understand the predictions made. Two explainability techniques were applied: an ante-hoc approach with an Attention layer and a post-hoc approach using Shapley Additive Explanations (SHAP). The Attention layer identified carbon monoxide (CO) and NO \({_2}\) as the most powerful features, while the SHAP analysis highlighted NO \({_2}\) as the predominant contributor to the predictions, followed by particulate matter (PM \(_{2.5}\) ) and CO. The results demonstrated that the target value significantly impacts the model’s forecast for both XAI techniques.