<i>FS</i> _SHAP: Global Interpretable Feature Selection Using XAI for the Classification of Spam SMS in Dravidian Languages

Recognizing the urgency of addressing the issue of deceptive fraudulent messages in the contemporary digital landscape that lead to financial losses or distressing incidents, detecting and identifying such fraudulent or spam messages is imperative before they reach users’ inboxes. This article proposes the eXplainable Artificial Intelligence (XAI) - SHAP-based feature selection framework (FSSHAP) for the classification of spam messages using the random forest algorithm. The model relies on the selection of significant handcrafted features based on SHapley Additive exPlanations (SHAP) values. The features are selected based on the threshold value determined using SHAP values. The research uses three datasets, including the short message services in Dravidian languages like Tamil, Kannada, Telugu, and Malayalam. The traditional feature selection techniques are compared with the proposed FSSHAP values approach in various performance metrics. This comprehensive approach aims to mitigate the impact of fraudulent messages, safeguarding users from potential harm in the digital realm.