Airport restroom cleanliness prediction using real time user feedback data

Large airports aim to offer a maximized experience to its passengers. A main contributor to customer experience is the cleanliness of restrooms, which is measured by feedback devices installed in restrooms at airports. This paper reviews to what extent real-time feedback data and classification techniques can be useful in practice to predict the cleanliness of restrooms. Within this topic, different class definitions of clean and unclean are introduced and a distinction is made between a combined prediction model that includes the entire environment and restroom specific prediction models that focus only on a single restroom. The dataset is imbalanced and visualizations show that there is class overlap. To overcame these limitations various sampling methods with two different encoding mechanisms are investigated. Sampling methods do not improve the performance of the combined prediction model but do improve the performance of some of the restroom-specific prediction models, especially those with a high class imbalance. The major cause of the unsatisfying performance is not class imbalance, but the data ambiguity that leads to class overlap. To obtain prediction models that are useful in practice, we provide recommendations regarding the dataset and how this should be enriched with features that are capable of distinguishing the two classes more clearly.