<p>Fraudulent claims in the automobile insurance industry pose a significant challenge, leading to substantial financial losses and necessitating the development of effective detection methods. This study aims to address this issue by proposing a Soft Voting ensemble learning approach to accurately detect fraudulent claims, particularly in imbalanced datasets. The proposed method combines predictions from multiple machine learning algorithms to enhance overall performance and robustness. To address the underrepresentation of fraudulent cases, we employ the SMOTE-NC oversampling technique during preprocessing. In addition, we use the Boruta algorithm and the Chi-square test to identify the most relevant factors influencing fraud detection. To ensure interpretability, we investigate Permutation Features Importance, SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP) and Counterfactual Explanations methods to improve transparency of machine learning models. The methodology is evaluated on a real-world automobile insurance dataset, where the proposed ensemble model outperforms individual classifiers. This study underscores the importance of integrating advanced machine learning techniques to improve fraud detection, while also promoting transparency and trust in automated decision-making through interpretability tools.</p>