Predicting Parking Availability from Mobile Payment Transactions with Positive Unlabeled Learning

Abstract

Cruising for parking in city centers is a problem for many motorists and for communities that need to reduce emissions. A widespread provision of parking assistance to address this problem requires a scalable system to generate availability information. Existing approaches to estimate the availability of parking spaces use supervised learning and depend on ground-truth labeling processes involving sensors or manual data collection. This dependency constraints the widespread roll-out and operation of such systems as the ground-truth data collection for model training, monitoring and retraining is prohibitively expensive. We describe a parking availability prediction system for paid on-street parking zones that does not depend on costly ground-truth labeling. The new approach uses solely data from parking ticket bookings via a mobile phone app. Every parking transaction serves as an implicit signal for the availability of one parking spot shortly before the booking. The system leverages this weak supervision signal by applying algorithms and metrics for positive-unlabeled learning (PU-learning). This approach enables the deployment in diverse regions, as well as the scalable monitoring and retraining of models. We evaluate our framework on a public dataset from Seattle.