Unsupervised text encoding models have recently fueled substantial progress\nin NLP. The key idea is to use neural networks to convert words in texts to\nvector space representations based on word positions in a sentence and their\ncontexts, which are suitable for end-to-end training of downstream tasks. We\nsee a strikingly similar situation in spatial analysis, which focuses on\nincorporating both absolute positions and spatial contexts of geographic\nobjects such as POIs into models. A general-purpose representation model for\nspace is valuable for a multitude of tasks. However, no such general model\nexists to date beyond simply applying discretization or feed-forward nets to\ncoordinates, and little effort has been put into jointly modeling distributions\nwith vastly different characteristics, which commonly emerges from GIS data.\nMeanwhile, Nobel Prize-winning Neuroscience research shows that grid cells in\nmammals provide a multi-scale periodic representation that functions as a\nmetric for location encoding and is critical for recognizing places and for\npath-integration. Therefore, we propose a representation learning model called\nSpace2Vec to encode the absolute positions and spatial relationships of places.\nWe conduct experiments on two real-world geographic data for two different\ntasks: 1) predicting types of POIs given their positions and context, 2) image\nclassification leveraging their geo-locations. Results show that because of its\nmulti-scale representations, Space2Vec outperforms well-established ML\napproaches such as RBF kernels, multi-layer feed-forward nets, and tile\nembedding approaches for location modeling and image classification tasks.\nDetailed analysis shows that all baselines can at most well handle distribution\nat one scale but show poor performances in other scales. In contrast,\nSpace2Vec's multi-scale representation can handle distributions at different\nscales.\n