🌇GURPP is a Graph-based Urban Region Pre-training and Prompting framework for region representation learning. Specifically, we first construct an urban region graph that integrates detailed spatial entity data for more effective urban region representation. Then, we develop a subgraph-centric urban region pre-training model to capture the heterogeneous and transferable patterns of interactions among entities. To further enhance the adaptability of these embeddings to different tasks, we design two graph-based prompting methods to incorporate explicit/hidden task knowledge. Extensive experiments on various urban region prediction tasks and different cities demonstrate the superior performance of our GURPP framework. This repository hosts the code of GURPP.

The Overview of GURPP is shown as follows:

1. Create urban region graph that integrates specific urban entities.
2. Extract subgraph of each region according to the graph pattern.
3. Use multi-view learning to learn the subgraph representations with GURP model.
4. Adapt the region embeddings with manually-designed prompt method and task-learnable prompt learning method.

• python==3.8.8
• pytorch==1.13.1
• numpy==1.23.5
• dgl==2.2.1
• scikit-learn==1.2.0
• pandas==1.5.2

+---GURPP
|   |   downstream_task.py
|   |   emb.npy
|   |   gurpp_args.py
|   |   load_graph_data.py
|   |   loss.py
|   |   README.md
|   |   test_gurp.py
|   |   train_gurp.py
|   |   train_learnable_prompt.py
|   |
|   +---data
|   |   +---nymhtkg
|   |   |   |   flow_in_per_hour.npy
|   |   |   |   flow_out_per_hour.npy
|   |   |   |   mobility_distribution.npy
|   |   |   |   region_si_img.npy
|   |   |   |
|   |   |   \---kge_pretrained_transR
|   |   |       \---TransR_UrbanKG_1
|   |   \---task
|   |           carbon_counts.npy
|   |           check_counts.npy
|   |           crime_counts.npy
|   |           income_counts.npy
|   |
|   +---experiments
|   |   \---gurp_model
|   |   \---gurp_prompt
|   |
|   +---figure
|   \---model
|           gurp.py
|           hgt.py
|           prompt.py

To reproduce the GURP results, you can run the following command:

python test_gurp.py

The test log will be saved in the experiments/gurp_prompt directory. Please ensure that this folder has been created following code structure before testing, otherwise, you might encounter a 'file not found' error.

To pre-train the GURP model, please following the steps below:

1. Download the pre-trained kge embeddings from UrbanKG_TransR_entity. After downloading, move the dataset to the data/nymhtkg/kge_pretrained_transR/TransR_UrbanKG_1 directory.

• mkdir -p data/nymhtkg/kge_pretrained_transR/TransR_UrbanKG_1

2. Download the prepared NYC Manhattan KG from mht180_prepared. After downloading, move the dataset to the data/nymhtkg directory.

• mkdir -p data/nymhtkg

3. You then can run the following command to pre-train the GURP model.

• python train_gurp.py

4. The training log will be saved in the experiments/gurp_model directory. Please ensure that this folder has been created following code structure before training, otherwise, you might encounter a 'file not found' error.

To train task-learnable prompt model, you can run the following command.

python train_learnable_prompt.py

And the training log will be saved in the experiments/gurp_prompt directory. Please ensure that this folder has been created following code structure before training, otherwise, you might encounter a 'file not found' error.

Manually-designed prompt can be designed by following steps.

1. Set the construct parameter if_test of GURPData as true.

2. Modify the method get_region_sub_test_all() in load_graph_data.py and then input them to the pretrained model to infer the results.

# fanout configuration to control the inclusion or exclusion of specific data types
# 0: exclude (ignore) the information
# -1: include the information
fanout = {'JCateOf': 0, 'BrandOf': 0, 'Cate1Of': 0, 'HasJunc': -1, 'HasPoi': -1, 'HasRoad': -1,
            'NearBy': -1, 'RCateOf': 0}
sub_graph = dgl.sampling.sample_neighbors(self.hg, {'region': i}, fanout,
                                            edge_dir='out', copy_ndata=True, copy_edata=True)

sub_poi_nodes = sub_graph.edges(etype='HasPoi')[1]
sub_road_nodes = sub_graph.edges(etype='HasRoad')[1]
sub_junc_nodes = sub_graph.edges(etype='HasJunc')[1]
sub_region_nodes = sub_graph.edges(etype='NearBy')[1]

seed = 0
random.seed(seed)
# Randomly sample 90% of the POI nodes, road nodes and junction nodes
sub_poi_nodes = random.sample(sub_poi_nodes.tolist(), int(len(sub_poi_nodes) * 0.9))
sub_road_nodes = random.sample(sub_road_nodes.tolist(), int(len(sub_road_nodes) * 0.9))
sub_junc_nodes = random.sample(sub_junc_nodes.tolist(), int(len(sub_junc_nodes) * 0.9))

sub_poi_nodes = torch.tensor(sub_poi_nodes, dtype=torch.long).to(self.graph_device)
sub_road_nodes = torch.tensor(sub_road_nodes, dtype=torch.long).to(self.graph_device)
sub_junc_nodes = torch.tensor(sub_junc_nodes, dtype=torch.long).to(self.graph_device)

# Sample neighbors for POI brand, POI category, junction category and road category
# 0: exclude (ignore) the information
# -1: include the information
poi_brand_graph = dgl.sampling.sample_neighbors(self.hg, {'poi': sub_poi_nodes},
                                                {'JCateOf': 0, 'BrandOf': -1, 'Cate1Of': 0, 'HasJunc': 0,
                                                    'HasPoi': 0, 'HasRoad': 0, 'NearBy': 0, 'RCateOf': 0},
                                                edge_dir='out', copy_ndata=True, copy_edata=True)
poi_cate1_graph = dgl.sampling.sample_neighbors(self.hg, {'poi': sub_poi_nodes},
                                                {'JCateOf': 0, 'BrandOf': 0, 'Cate1Of': -1, 'HasJunc': 0,
                                                    'HasPoi': 0, 'HasRoad': 0, 'NearBy': 0, 'RCateOf': 0},
                                                edge_dir='out', copy_ndata=True, copy_edata=True)
junc_cate_graph = dgl.sampling.sample_neighbors(self.hg, {'junction': sub_junc_nodes},
                                                {'JCateOf': -1, 'BrandOf': 0, 'Cate1Of': 0, 'HasJunc': 0,
                                                    'HasPoi': 0, 'HasRoad': 0, 'NearBy': 0, 'RCateOf': 0},
                                                edge_dir='out', copy_ndata=True, copy_edata=True)
road_cate_graph = dgl.sampling.sample_neighbors(self.hg, {'road': sub_road_nodes},
                                                {'JCateOf': 0, 'BrandOf': 0, 'Cate1Of': 0, 'HasJunc': 0,
                                                    'HasPoi': 0, 'HasRoad': 0, 'NearBy': 0, 'RCateOf': -1},
                                                edge_dir='out', copy_ndata=True, copy_edata=True)

• Urban Region Graph in Manhattan of NYC (in the format of triples):
  • mht180_prepared.csv: The urban region graph in Manhattan of NYC.
• Pretrained nymht KGE embeddings for urban region entities:
  • TransR_UrbanKG_1: The pretrained nymht KGE embeddings for urban region entities.