This is the Pytorch implementation of our ICDE'23 paper: A Contextual Master-Slave Framework on Urban Region Graph for Urban Village Detection.

• python 3.x
• torch == 1.7.1
• dgl == 0.6.1

The three real-world datasets used in our paper can be downloaded here:

Put all data files in dir: data/

(1) city_poi_features.npy: (num_nodes, 64) numpy array of our constructed POI features.

• poi_features[:, 0] : region id (not used)
• poi_features[:, 1] : Index of basic living facility
• poi_features[:, 2:26] : Category distribution in the given region
• poi_features[:, 26:50] : Category distribution in the 3x3 grids centered by the given region
• poi_features[:, 50:] : POI radius features

(2) city_img_features.npy: (num_nodes, 4096) numpy array of image features extracted by the VGG16 model pre-trained on ImageNet.

(3) city_labels.npy: (num_nodes, ) numpy array of ground truth binary label indicating that whether a region is urban village (y=1) or not (y=0).

• Notes: The y value for unlabeled regions are also set to 0 in this array, they will be masked when calculating the loss and evaluation metrics.

(4) city_adj.npy: (2, num_edges) array of edges of the urban region graph (URG).

• adj[0]: start nodes list of edegs
• adj[1]: end nodes list of edegs

(5) city_mask.json : the dict recording the region id in train / val / test set under each dataset split for nested cross validation.

Set the hyper-parameters in script.sh, and train the master model by:

sh script.sh master train

Then, train the slave model based on the saved master model by:

sh script.sh slave train

Make urban village prediction with the trained slave model by:

sh script.sh slave test

Please refer to the full version:

Please cite our work if you find our data/code/paper is useful to your work: