Exploring how independent variables influence parking occupancy prediction: toward a model results explanation with SHAP values

• Published in: Progress in artificial intelligence Vol. 11; no. 4; pp. 367 - 396
• Main Authors: Errousso, Hanae, Abdellaoui Alaoui, El Arbi, Benhadou, Siham, Medromi, Hicham
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.12.2022; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Automobile parking; City centres; Computational Intelligence; Computer Imaging; Computer Science; Computing time; Control; Data Mining and Knowledge Discovery; Independent variables; Machine learning; Mechatronics; Natural Language Processing (NLP); Parking; Pattern Recognition and Graphics; Real time; Regular Paper; Robotics; Vision; Explanation model; Parking occupancy; Game theory; Machine learning methods; SHAP values; Prediction
• ISSN: 2192-6352; 2192-6360
• DOI: 10.1007/s13748-022-00291-5

Finding a parking space is a difficult challenge that drivers face on a daily basis in urban neighborhoods around the world. They often report that desirable spaces near to their destination are either unavailable or very expensive, extending further the search time and congesting even more city centers. Intelligent parking solutions can integrally solve this ongoing problem by better managing existing resources. They allow drivers to access real-time information on parking space availability, collected with different detection techniques (crowdsourcing, parking meters, sensors). Some of these systems also encompass opportunistic services, such as forecasting, needed to adapt to unforeseen dynamic situations. Hence, we presented, in this paper, a methodology for predicting car park occupancy rates using four different machine learning algorithms. Each of these methods is trained with four feature sets to exemplify how information quality impacts prediction accuracy. In addition to achieving high accuracy, it is absolutely crucial to interpret model outputs and analyze each individual feature’s importance. That's why we developed an explanation model based on SHAP values. We implemented our proposal exploiting five months of real-time parking data broadcast by Aarhus City Council. Results show that the best-obtained predictions are by far very accurate with a coefficient of determination ( R 2 ) that achieves 0.988 and a mean absolute error that doesn't exceed 2.021%, while requiring a very low computing time that is only 5 s.