Interpretable Machine Learning for Undeclared Work Prediction

Machine learning models have vastly proved their contribution to decision-making in many aspects of daily life. Yet, their acceptance is confined, especially in public authorities, due to the lack of trust and reliability in their results, often credited to the shortage of “black box” models' explainability and interpretability. The present study illustrates an application of interpretable machine learning to predict undeclared work and other labour law infringements, using real-life data of past inspections and employment declarations obtained from a Labour Inspectorate. Undeclared work is a severe socioeconomic issue causing significant losses in taxes and social security contributions and undermining competitiveness and people and state welfare, thus rendering its detection and tackling a critical goal for the enforcement authorities in charge. We employ Lazy Associative Classification and encompass data sampling methods to deal with class overlapping and imbalance problems, achieving increased prediction performance, tripling the inspection yields with regard to undeclared work detection, and, in parallel, transparency, offering precise explanations for each prediction, as well as general insights in the prevalence and feature patterns of undeclared work. This research also discusses the multiple benefits of embodying this machine learning approach in an inspection recommendation system for responsible enforcement authorities, with the most significant being simplicity, flexibility, broadened adaptability to the users' diverse and ever-changing needs, and enhanced acceptance and confidence of the users in the system inspection suggestions.