Travel behaviour Data Fusion Model based on APP-based survey and floating phone data

Floating Phone Data (FPD), derived from mobile phone signaling events, is widely available and valuable for mobility monitoring. With its large sample size, FPD has proven to be a reliable data source for origin-destination relations; however, transport mode detection remains a significant challenge. While its big data attributes make it suitable for machine learning approaches, the lack of labeling in FPD complicates the straightforward application of such models. Previous studies have developed machine learning models to infer transport modes for FPD, but these often require targeted surveys where FPD and travel diaries are recorded together in large samples, making them less practical for widespread applications. In this study, we introduce a method to build a Random Forest-based transport mode prediction model for FPD that does not require FPD for training. Instead, we use GPS-supported, app-based travel diaries from a panel survey conducted in the urban development area aspern Seestadt in Vienna. The model predicts nine transport modes: Bicycle, Car, Taxi/Uber, Bus, Subway, Train, Tramway, Walking, and Other. To enhance performance, we introduced key cross-sections around the Seestadt area as spatial features, creating a cordon-like setting, and mined additional features for fine-grained public transport mode detection from General Transit Feed Specification (GTFS) data. To validate the model, we conducted an independent survey, treated as ground truth, collecting both FPD and travel diary data simultaneously. Our results highlight the effectiveness of cross-sections as spatial features for mode detection models. The final model achieved an overall accuracy of 0.82 on FPD, with strong performance for Subway, Train, and Tram modes. However, the model struggled with the Walking mode due to limited training data—a general limitation of this study. We aim to significantly expand the sample size in further research to enable more robust evaluations. Overall, our approach makes a valuable contribution to the practical application of transport mode detection models for FPD.