Improving the performance of machine learning algorithms using conceptual models: a case study of auto insurance

Abstract

The integration of domain knowledge into machine learning models has been proposed as a means to address the limitations of purely data-driven approaches. Traditional machine learning techniques often rely on pre-defined, fixed data structures, which can overlook valuable context-specific insights that domain knowledge provides. This study investigates the impact of incorporating domain knowledge into the preprocessing and feature engineering stages of machine learning models, specifically focusing on decision tree algorithms and Support Vector Machines (SVM) within the insurance sector. To evaluate the effectiveness of this integration, this study compares the performance of models trained on a pre-defined dataset (A) with models trained on the same dataset after it was enhanced with domain-specific knowledge (Revised A). The results demonstrate that the integration of domain-specific guidelines into the feature engineering process significantly improved the accuracy and reliability of the predictive models, particularly in complex scenarios such as predicting customer profitability. In scenarios where domain knowledge played a crucial role in refining features that capture relationships within the insurance data, the enhanced models outperformed the original ones. Conversely, for tasks where the domain knowledge had less influence, the performance improvement was marginal. These findings suggest that integrating domain knowledge into machine learning processes can provide a meaningful boost in model effectiveness, but the benefits are context-dependent.