Scholarlink

<p>In this study, we introduce a novel predictive approach rooted in the field of explainable AI, designed to provide sketches of the action of a non-interpretable machine learning model. The use case is psychological profiling and concerns the alignment between individuals’ personalities and specific job roles within the fields of psychology and career development. Our approach builds upon the foundation of BAPC (Before and After prediction Parameter Correction) (Sobieczky and Geiß in Explainable AI by BAPC—Before and After correction Parameter Comparison, 2023), an explainable AI technique designed to elucidate predictions from models with interpretable parameters of a surrogate ’base’ model. We extend this method by incorporating error amplification&#xa0;(EA), a new technique effectively integrating decision trees as our base model. In the three-step methodology, rooted in the principles of BAPC, enhanced by error amplification, we firstly train a base model using decision trees and identify the model’s errors. Second, we introduce an AI model to predict the errors made by the base model. Subsequently, we amplify these predicted errors in a localized manner around the individual point of interest. In the third step, we retrain the base model using the error-amplified data set. The disparity between the decision boundaries generated by the initial and the retrained decision trees provides explainable insights into enhancing specific personality features. While our method finds its application in personality-job alignment here, its versatility extends to a wide range of scenarios, offering valuable insights into data set feature interpretability by elucidating their effective changes.</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

Explainable classification by local error amplification

  • 期刊论文

  • 发表日期：

    2025年8月28日

查看全文
  • Erika Dudkin,
  • Florian Sobieczky

摘要

In this study, we introduce a novel predictive approach rooted in the field of explainable AI, designed to provide sketches of the action of a non-interpretable machine learning model. The use case is psychological profiling and concerns the alignment between individuals’ personalities and specific job roles within the fields of psychology and career development. Our approach builds upon the foundation of BAPC (Before and After prediction Parameter Correction) (Sobieczky and Geiß in Explainable AI by BAPC—Before and After correction Parameter Comparison, 2023), an explainable AI technique designed to elucidate predictions from models with interpretable parameters of a surrogate ’base’ model. We extend this method by incorporating error amplification (EA), a new technique effectively integrating decision trees as our base model. In the three-step methodology, rooted in the principles of BAPC, enhanced by error amplification, we firstly train a base model using decision trees and identify the model’s errors. Second, we introduce an AI model to predict the errors made by the base model. Subsequently, we amplify these predicted errors in a localized manner around the individual point of interest. In the third step, we retrain the base model using the error-amplified data set. The disparity between the decision boundaries generated by the initial and the retrained decision trees provides explainable insights into enhancing specific personality features. While our method finds its application in personality-job alignment here, its versatility extends to a wide range of scenarios, offering valuable insights into data set feature interpretability by elucidating their effective changes.