Activity type detection of mobile phone data based on self-training: Application of the teacher–student cycling model

Incorporating mobile phone data, known for its high spatial and temporal resolution and extensive population coverage, into Activity-Based Models (ABM) for understanding individual travel and activity behaviors is a current research hotspot. However, applying them in ABM building is not straightforward because they miss key information — activity type. In this paper, we present an activity types detection method named the Teacher–Student Cycling model based on a self-training framework. Our model can combine travel survey data with extensive prior knowledge and mobile phone data. We introduce two different resolutions of mobile phone datasets to test the model performance. Our results show that our proposed model can achieve good performance on datasets of different resolutions. Our model, one of the semi-supervised learning models that uses a mixture of labeled and unlabeled knowledge performs better than the supervised learning model that uses labeled knowledge alone. In addition, our model improves the overall detection accuracy by 7% over the second-best semi-supervised learning model and improves the detection accuracy of secondary activities by up to 17%. Our model can be valuable in generating daily activity schedules for agent-based models.