Predicting Urban Problems: A Comparison of Graph-based and Image-based Methods

Local governments have social problems closely related to urban spatial features and human behaviors, such as graffiti, littering, and illegally parked bicycles. Our goal is to estimate the current situation of these urban problems. In this paper, we focused on the littering problem as one of the urban problems, and we compared a graph-based method using knowledge graph embedding and an image-based method using convolutional neural networks (CNN). Specifically, in the graphbased method, we first design and construct the knowledge graph based on the results of the airflow simulation and geospatial features in urban areas. Then, we generate the vector data from the knowledge graph using RDF2vec, and estimate the litter distributions using the vector data. As the result, both of the two methods achieved high and approximately same accuracy, although the graph-based method has a smaller amount of numerical information than the image-based method.