Delineating urban functional use from Points of Interest data with Doc2Vec model

Understanding how human activities and the uses of buildings are distributed and change (i.e. urban functional use) helps to detect urban problems, evaluate planning strategies, support policymaking and modelling. With precise geolocation and detailed classification indicating preference of urban activities, ’Points of Interest’ type of data that tends to identify where certain uses are (e.g. restaurants, supermarkets) have been widely used in delineating urban functional area. However, previous works with frequency-based approach (i.e. TF-IDF and LDA) or neural word embedding (i.e. Word2Vec) don’t consider spatial context around POIs and therefore the spatial disparity of POIs is absent of when identifying urban functional areas. This paper introduces Doc2Vec to include the local geographic characteristics into neural word embedding, exploring the relationship between fine-scale POI classes, and identifying functional areas.