Scholarlink

The widespread growth of social media has intensified the presence of fake profiles, often exploited to spread misinformation, conduct scams, and manipulate public opinion. This study introduces an explainable hybrid model that integrates eXtreme Gradient Boosting (XGBoost) and Deep Neural Networks (DNN), enhanced through a novel SHAP (SHapley Additive Explanations)-based feature engineering strategy. SHAP values, computed from the trained XGBoost model, quantify feature influence and are fused with standardized features to create enriched representations for DNN training. Unlike conventional approaches that employ SHAP solely for post-hoc interpretability, our method leverages SHAP during feature engineering to improve predictive performance. For interpretability at inference time, LIME (Local Interpretable Model-Agnostic Explanations) provides instance-level insights for each prediction. The proposed system is evaluated on a balanced dataset of 5,000 Instagram profiles and benchmarked against individual classifiers (e.g., Random Forest, SVM) and ensemble techniques. Achieving 99.1% accuracy, the model outperforms all baselines. Finally, the system is deployed to support both manual input and automated data retrieval, bridging performance with interpretability and contributing a reliable solution for social media integrity.

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

SHAP for Explainable Feature Engineering and LIME-Based Interpretability to Enhance Fake Profile Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • V. Praveen Goud,
  • Raghava Morusupalli

摘要

The widespread growth of social media has intensified the presence of fake profiles, often exploited to spread misinformation, conduct scams, and manipulate public opinion. This study introduces an explainable hybrid model that integrates eXtreme Gradient Boosting (XGBoost) and Deep Neural Networks (DNN), enhanced through a novel SHAP (SHapley Additive Explanations)-based feature engineering strategy. SHAP values, computed from the trained XGBoost model, quantify feature influence and are fused with standardized features to create enriched representations for DNN training. Unlike conventional approaches that employ SHAP solely for post-hoc interpretability, our method leverages SHAP during feature engineering to improve predictive performance. For interpretability at inference time, LIME (Local Interpretable Model-Agnostic Explanations) provides instance-level insights for each prediction. The proposed system is evaluated on a balanced dataset of 5,000 Instagram profiles and benchmarked against individual classifiers (e.g., Random Forest, SVM) and ensemble techniques. Achieving 99.1% accuracy, the model outperforms all baselines. Finally, the system is deployed to support both manual input and automated data retrieval, bridging performance with interpretability and contributing a reliable solution for social media integrity.