No abstract

Existing studies on activity location recognition based on mobile phone data has made great progresses. However, current studies generally assume constant distance threshold when performing activity location clustering, and ignore the influence of base station layout on positioning accuracies of mobile phone data. Given different recognition accuracy requirements, the authors propose two methods to recognise activity locations: (1) An improved hierarchical agglomerative clustering algorithm that integrates a genetic algorithm component to search and dynamically adjust optimal distance thresholds based on base station densities; (2) The recognition method based on Bi‐directional long short‐term memory network that classifies travel statuses of mobile phone traces. Results show that, compared with existing methods, the activity location recognition accuracy of the proposed hierarchical agglomerative clustering algorithm increases by about 5%. The Bi‐directional long short‐term memory network model further outperforms the improved hierarchical agglomerative clustering, especially in the aspect of recognising non‐commuting activity locations. However, the Bi‐directional long short‐term memory network model training requires the users’ actual travel information, so there are certain obstacles in popularising Bi‐directional long short‐term memory network in practice.

Trip end identification based on mobile phone data has been widely investigated in recent years. However, the existing studies generally use fixed clustering radii (CR) in trip end clustering algorithms, but ignore the influence of base station (BS) densities on the positioning accuracy of mobile phone data. This paper proposes a new two-step method for identifying trip ends: (1) Genetic Algorithm (GA) is utilized to optimize the CRs of DBSCAN under different BS densities. (2) We propose an improved Fast-DBSCAN (F-DBSCAN) for two objectives. One is for improving identification accuracies; the parameter CRs for judging core points can be dynamically adjusted based on the BS density around each mobile phone trace. The other is for reducing time complexity; a fast clustering improvement for the algorithm is proposed. Mobile phone data was collected by real-name volunteers with support from the communication operator. We compare the identification accuracy and time complexity of the proposed method with the existing ones. Results show that the accuracy is raised to 85%, which is approximately 6% higher than the existing methods. Meanwhile, the median running time can be reduced by about 76% by the fast clustering improvement. Especially for noncommuting trip ends, the identification accuracy can be increased by 8%. The average identification errors of travel time and trip end coordinates are reduced by about 12 min and 321 m, respectively.