Rastgele Orman'ın ötesinde: Yorumlanabilirlik gereksinimleri ile sigorta satın alma tahmini için gelişmiş topluluk ve hibrit modeller

This study focuses on creating machine learning models to predict whether people will buy insurance. We developed and tested sophisticated ensemble models that aim to be both highly accurate in their predictions and easy to understand: an important balance since insurance companies need to explain their decision-making processes to meet regulatory requirements. Using a dataset of 53,503 customer records, a comprehensive data preparation pipeline transformed 20 raw features into 100 engineered variables, thereby capturing temporal, behavioural, and categorical patterns. Class imbalance was addressed using Synthetic Minority Oversampling Technique (SMOTE) which ensures equitable model learning. Models which include Random Forest, Support Vector Machine (SVM), LightGBM, Graph Neural Networks (GNNs), and hybrid designs were compared, with Random Forest and LightGBM achieving near-perfect performance (F1-score: 1.0000 and 0.9999, respectively). GNNs, integrated by means of a k-NN similarity graph, showed scalability but small performance gains due to the sufficiency of original features. Since interpretability was a focus of this research, SHAP and LIME were used and their use revealed that temporal features such as "Days_Since_Purchase" and "Purchase_Year" were the main contributing factors. Also, optimization with multiple objectives using Optuna achieved a Pareto-optimal configuration (ROC-AUC: 1.0, F1-score: 0.9896, fidelity: 0.9843, sparsity: 3.0) which satisfies strict regulatory requirements (GDPR, NAIC). A case study on 1,000 customers showed the trade-off between high-performance LightGBM+GNN (ROC-AUC: 0.6211) and interpretable Decision Trees (ROC-AUC: 0.5380), with stakeholder feedback favouring transparency. The study recommends making temporal and monetary features priorities for customer segmentation and examining time-evolving GNN architectures and alternative graph constructions for future research to improve predictive power in noisier datasets. This framework provides a scalable, interpretable solution for insurance analytics, and that supports targeted marketing and compliance.