Loyal to your city? A data mining analysis of a public service loyalty program

• The visiting behavior of users of a government-issued loyalty card is analyzed.
• The visiting behavior of users is modeled in a spatial and a spatio-temporal fashion.
• Naive Bayes classifier and Support Vector Machine are trained for predictive models.
• Using temporal data down to the most fine-grained level leads to better predictions.
• We report promising results regarding location, loyalty and defect prediction.

Customer loyalty programs are largely present in the private sector and have been elaborately studied. Applications from the private sector have found resonance in a public setting, however, simply extrapolating research results is not acceptable, as their rationale inherently differs. This study focuses on data from a loyalty program issued by the city of Antwerp (Belgium). The aim of the loyalty card entails large citizen participation, however, an active user base of only 20 % is reached. Predictive techniques are employed to increase this number. Using spatial behavioral user information, a Naive Bayes classifier and a Support Vector Machine are used which result in models capable of predicting whether a user will actively use its card, whether a user will defect in the near future and which locations a user will visit. Also, a projection of spatial behavioral data onto even more fine-grained spatio-temporal data is performed. The results are promising: the best model achieves an AUC value of 92.5 %, 85.5 % and 88.12 % (averaged over five locations) for the predictions, respectively. Moreover, as behavior is modeled in more detail, better predictions are made. Two main contributions are made in this study. First, as a theoretical contribution, fine-grained behavioral data contributes to a more sound decision-making process. Second, as a practical contribution, the city of Antwerp can now make tailored strategic decisions to increase its active user base.