Feature-Driven vs Language-Based AI Online Gambling Addiction Modeling: Exploring Interpretability through XGBoost and LLMBased RAG

Abstract

The rise of online gambling has increased concern around identifying behavioral
addiction in digital environments. Current predictive systems offer limited
interpretability and justification for individual-level risk assessments as they
often operate as black boxes. This study proposes a hybrid framework that
combines a popular machine learning model (XGBoost) with a language-based
Retrieval-Augmented Generation (RAG) system to address the current
challenges. A combination of user-level behavioral and demographic data was
used as input for a trained XGBoost classifier and SHAP (SHapley Additive
exPlanations) was also applied to find which features contribute the most to
addiction after evaluation. These insights were then incorporated into a Large
Language Model (LLM)-based RAG pipeline using sentence-transformer
embeddings and FAISS vector retrieval to generate individualized text
justifications for each user classification. Through label refinement based on
SHAP-ranked feature thresholds and targeted model tuning, the system achieved
improved generalization and classification stability, resulting in an AUC of
0.87 while preserving clear, human-readable explanations via the RAG pipeline.
This approach demonstrates the potential of integrating structured and
unstructured AI techniques in addiction research and risk screening to support
more accountable and understandable behavioral health interventions.