Identifying key grid cells for crowd flow predictions based on CNN-based models with the Grad-CAM kit

Scholars have long sought to identify key city locations that have a pronounced effect on the flow of people under various conditions. Identifying key locations makes it possible for government and/or enterprises to obtain accurate estimates as to the flow of people and thereby formulate reasonable management strategies. Note however that much of the previous research based on professional knowledge or machine learning fails to provide accurate results. Some researchers have employed CNN-based models to predict the flow of people, claiming that those models can extract freatures from multiple locations to improve prediction accuracy. Theoretically, the features extracted using CNN-based models could be used themselves as key locations for specific problems; however, the fact that the features are consolidated via multiple operations often renders interpretation difficult when applied to a real-world setting. In the current study, we developed a novel approach to the identification of key locations based on the results of a CNN-based model and the Grad-Cam kit. The structure of a CNN-based models can have a profound effect on the results of the Grad-Cam kit; therefore, we compared the results obtained using three state-of-art models as well as our CNN-LSTM. Actual flow patterns based on telecommunication data in Taipei were used to verify the efficacy of the proposed method.