{"id":97157,"date":"2026-06-13T17:57:52","date_gmt":"2026-06-13T17:57:52","guid":{"rendered":"https:\/\/youzum.net\/a-coding-implementation-on-spatial-graph-neural-networks-for-urban-function-inference-using-city2graph-osmnx-and-pytorch-geometric\/"},"modified":"2026-06-13T17:57:52","modified_gmt":"2026-06-13T17:57:52","slug":"a-coding-implementation-on-spatial-graph-neural-networks-for-urban-function-inference-using-city2graph-osmnx-and-pytorch-geometric","status":"publish","type":"post","link":"https:\/\/youzum.net\/fr\/a-coding-implementation-on-spatial-graph-neural-networks-for-urban-function-inference-using-city2graph-osmnx-and-pytorch-geometric\/","title":{"rendered":"A Coding Implementation on Spatial Graph Neural Networks for Urban Function Inference Using city2graph, OSMnx, and PyTorch Geometric"},"content":{"rendered":"

In this tutorial, we build an end-to-end spatial graph learning pipeline using city2graph<\/strong><\/a>. We start by collecting real urban POI data and street network information from OpenStreetMap, with a synthetic fallback to ensure the workflow remains reliable. We then engineer spatial features, construct multiple proximity graph families, and compare how different graph-building strategies represent the same urban environment. After that, we create both heterogeneous and homogeneous graph structures, convert them into PyTorch Geometric format, and train a GraphSAGE model to predict POI categories from spatial structure. Through this process, we integrate geospatial data processing, graph construction, and GNN-based urban function inference into a single practical workflow.<\/p>\n

Installing city2graph and Importing Geospatial and Graph Learning Libraries<\/strong><\/h3>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
!pip -q install \"city2graph[cpu]\" osmnx contextily scikit-learn 2>\/dev\/null\nimport warnings, numpy as np, pandas as pd, geopandas as gpd\nwarnings.filterwarnings(\"ignore\")\nfrom shapely.geometry import Point\nimport matplotlib.pyplot as plt\nimport city2graph as c2g\nprint(\"city2graph version:\", getattr(c2g, \"__version__\", \"unknown\"))\nprint(\"PyTorch \/ PyG available:\", c2g.is_torch_available())\nimport torch\nimport torch.nn.functional as F\nfrom torch_geometric.nn import SAGEConv, to_hetero\nfrom torch_geometric.utils import to_undirected\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.neighbors import NearestNeighbors\nfrom sklearn.metrics import accuracy_score, f1_score\nfrom sklearn.decomposition import PCA\nSEED = 42\nnp.random.seed(SEED); torch.manual_seed(SEED)\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We begin by installing the required libraries and importing the geospatial, graph learning, and machine learning tools used throughout the tutorial. We verify that city2graph and PyTorch Geometric are available so the rest of the workflow can run properly. We also set a fixed random seed to make the graph construction, training split, and model results more reproducible.<\/p>\n
Collecting OpenStreetMap POI Data with a Synthetic Fallback<\/strong><\/h3>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
CENTER = (35.6595, 139.7005)\nDIST_M = 1100\nTAG_QUERIES = {\n   \"food\":      {\"amenity\": [\"restaurant\", \"cafe\", \"fast_food\", \"bar\", \"pub\"]},\n   \"retail\":    {\"shop\": True},\n   \"education\": {\"amenity\": [\"school\", \"university\", \"college\", \"kindergarten\", \"library\"]},\n   \"health\":    {\"amenity\": [\"hospital\", \"clinic\", \"pharmacy\", \"doctors\", \"dentist\"]},\n}\ndef to_points(gdf):\n   g = gdf.copy()\n   g[\"geometry\"] = g.geometry.representative_point()\n   return g\npoi_gdf, segments_gdf = None, None\ntry:\n   import osmnx as ox\n   ox.settings.use_cache = True\n   ox.settings.log_console = False\n   frames = []\n   for label, tags in TAG_QUERIES.items():\n       try:\n           f = ox.features_from_point(CENTER, tags=tags, dist=DIST_M)\n           f = f[f.geometry.notna()]\n           if len(f):\n               f = to_points(f)[[\"geometry\"]].copy()\n               f[\"category\"] = label\n               frames.append(f)\n       except Exception as e:\n           print(f\"  (skip {label}: {e})\")\n   if not frames:\n       raise RuntimeError(\"No POIs returned from Overpass.\")\n   poi_gdf = gpd.GeoDataFrame(pd.concat(frames, ignore_index=True), crs=\"EPSG:4326\")\n   G = ox.graph_from_point(CENTER, dist=DIST_M, network_type=\"walk\")\n   segments_gdf = ox.graph_to_gdfs(G, nodes=False, edges=True).reset_index(drop=True)[[\"geometry\"]]\n   print(f\"OSM acquisition OK -> {len(poi_gdf)} POIs, {len(segments_gdf)} street segments\")\nexcept Exception as e:\n   print(f\"OSM unavailable ({e}) -> generating synthetic clustered POIs.\")\n   rng = np.random.default_rng(SEED)\n   cats = list(TAG_QUERIES.keys())\n   centers = rng.uniform(-0.01, 0.01, size=(8, 2)) + np.array(CENTER[::-1])\n   rows = []\n   for ci, c in enumerate(centers):\n       dom = cats[ci % len(cats)]\n       n = rng.integers(40, 90)\n       pts = c + rng.normal(0, 0.0016, size=(n, 2))\n       for (lon, lat) in pts:\n           cat = dom if rng.random() < 0.75 else rng.choice(cats)\n           rows.append({\"geometry\": Point(lon, lat), \"category\": cat})\n   poi_gdf = gpd.GeoDataFrame(rows, crs=\"EPSG:4326\")\n   segments_gdf = None\n   print(f\"Synthetic dataset -> {len(poi_gdf)} POIs\")\nif len(poi_gdf) > 700:\n   poi_gdf = poi_gdf.sample(700, random_state=SEED).reset_index(drop=True)\nmetric_crs = poi_gdf.estimate_utm_crs()\npoi_gdf = poi_gdf.to_crs(metric_crs).reset_index(drop=True)\nif segments_gdf is not None:\n   segments_gdf = segments_gdf.to_crs(metric_crs)\nprint(\"Class balance:n\", poi_gdf[\"category\"].value_counts())\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We collect real POI data from OpenStreetMap around Shibuya, Tokyo, and group the locations into broad urban function categories such as food, retail, education, and health. We also download the walkable street network so that the POIs can later be connected with urban-form features. If the OSM request fails, we generate a synthetic clustered dataset, which keeps the tutorial runnable even when online data access is unavailable.<\/p>\n
Engineering Spatial Features and Building Proximity Graph Families<\/strong><\/h3>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
poi_gdf[\"cx\"] = poi_gdf.geometry.x\npoi_gdf[\"cy\"] = poi_gdf.geometry.y\ncoords = poi_gdf[[\"cx\", \"cy\"]].to_numpy()\nnn = NearestNeighbors(radius=150.0).fit(coords)\npoi_gdf[\"local_density\"] = [len(idx) - 1 for idx in nn.radius_neighbors(coords, return_distance=False)]\nif segments_gdf is not None and len(segments_gdf):\n   try:\n       joined = gpd.sjoin_nearest(poi_gdf[[\"geometry\"]], segments_gdf[[\"geometry\"]],\n                                  distance_col=\"dist_street\")\n       poi_gdf[\"dist_street\"] = joined.groupby(level=0)[\"dist_street\"].min().reindex(poi_gdf.index).fillna(0.0)\n   except Exception:\n       poi_gdf[\"dist_street\"] = 0.0\nelse:\n   poi_gdf[\"dist_street\"] = 0.0\npoi_gdf[\"category\"] = poi_gdf[\"category\"].astype(\"category\")\npoi_gdf[\"label\"] = poi_gdf[\"category\"].cat.codes.astype(int)\nCLASS_NAMES = list(poi_gdf[\"category\"].cat.categories)\nprint(\"Classes:\", CLASS_NAMES)\ndef graph_stats(name, builder):\n   try:\n       nodes, edges = builder()\n       deg = pd.Series(np.r_[edges.index.get_level_values(0),\n                             edges.index.get_level_values(1)]).value_counts()\n       return name, len(edges), round(deg.mean(), 2), (nodes, edges)\n   except Exception as e:\n       return name, f\"ERR: {e}\", None, None\nbuilders = {\n   \"KNN (k=8)\":  lambda: c2g.knn_graph(poi_gdf, distance_metric=\"euclidean\", k=8, as_nx=False),\n   \"Delaunay\":   lambda: c2g.delaunay_graph(poi_gdf, as_nx=False),\n   \"Gabriel\":    lambda: c2g.gabriel_graph(poi_gdf, as_nx=False),\n   \"RNG\":        lambda: c2g.relative_neighborhood_graph(poi_gdf, as_nx=False),\n   \"EMST\":       lambda: c2g.euclidean_minimum_spanning_tree(poi_gdf, as_nx=False),\n   \"Waxman\":     lambda: c2g.waxman_graph(poi_gdf, distance_metric=\"euclidean\", r0=150, beta=0.6),\n}\nprint(\"n--- Proximity graph comparison ---\")\nprint(f\"{'graph':<14}{'#edges':>10}{'avg_degree':>12}\")\nbuilt = {}\nfor nm, b in builders.items():\n   name, ne, avgdeg, payload = graph_stats(nm, b)\n   print(f\"{name:<14}{str(ne):>10}{str(avgdeg):>12}\")\n   if payload: built[nm] = payload\nfig, axes = plt.subplots(1, 3, figsize=(16, 5))\nfor ax, key in zip(axes, [\"KNN (k=8)\", \"Delaunay\", \"EMST\"]):\n   if key in built:\n       n_, e_ = built[key]\n       e_.plot(ax=ax, linewidth=0.4, color=\"#3b7dd8\", alpha=0.6)\n       poi_gdf.plot(ax=ax, markersize=4, color=\"#d83b5c\")\n       ax.set_title(key); ax.set_axis_off()\nplt.suptitle(\"Spatial graph topologies on the same POI set\", y=1.02)\nplt.tight_layout(); plt.show()\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We engineer spatial features for each POI by extracting its projected coordinates, calculating local density, and estimating distance to the nearest street segment. We then assign category labels and build several families of proximity graphs, including KNN, Delaunay, Gabriel, RNG, EMST, and Waxman. We compare their edge counts and average degrees, then visualize selected graph topologies to see how differently they connect the same set of POIs.<\/p>\n
Constructing Heterogeneous and Homogeneous Graphs in PyTorch Geometric<\/strong><\/h3>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
nodes_dict = {}\nfor cat in CLASS_NAMES:\n   sub = poi_gdf[poi_gdf[\"category\"] == cat].copy().reset_index(drop=True)\n   nodes_dict[cat] = sub[[\"geometry\", \"cx\", \"cy\", \"local_density\"]]\ntry:\n   _, bridge_edges = c2g.bridge_nodes(nodes_dict, proximity_method=\"knn\", k=3,\n                                      distance_metric=\"euclidean\")\n   hetero = c2g.gdf_to_pyg(\n       nodes_dict, bridge_edges,\n       node_feature_cols={cat: [\"cx\", \"cy\", \"local_density\"] for cat in CLASS_NAMES},\n   )\n   print(\"nHeteroData node types:\", hetero.node_types)\n   print(\"HeteroData edge types:\")\n   for et in hetero.edge_types:\n       print(f\"   {et}: {hetero[et].edge_index.shape[1]} edges\")\nexcept Exception as e:\n   hetero = None\n   print(\"Heterogeneous build skipped:\", e)\nnodes, edges = c2g.knn_graph(poi_gdf, distance_metric=\"euclidean\", k=8, as_nx=False)\ndeg = pd.Series(np.r_[edges.index.get_level_values(0),\n                     edges.index.get_level_values(1)]).value_counts()\nnodes[\"degree\"] = deg.reindex(nodes.index).fillna(0).astype(float)\nfor col in [\"cx\", \"cy\", \"local_density\", \"dist_street\", \"label\"]:\n   if col not in nodes.columns:\n       nodes[col] = poi_gdf.loc[nodes.index, col].values\nFEATS = [\"cx\", \"cy\", \"local_density\", \"dist_street\", \"degree\"]\nnodes[FEATS] = StandardScaler().fit_transform(nodes[FEATS].astype(float))\ndata = c2g.gdf_to_pyg(nodes, edges, node_feature_cols=FEATS, node_label_cols=[\"label\"])\ndata.edge_index = to_undirected(data.edge_index)\ndata.x = data.x.float()\ny = data.y.long().view(-1)\nN, num_classes = data.num_nodes, int(y.max()) + 1\nprint(f\"nHomogeneous Data: {N} nodes, {data.edge_index.shape[1]} directed-edges, \"\n     f\"{data.x.shape[1]} features, {num_classes} classes\")\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We construct a heterogeneous multi-layer graph by separating POIs into node types based on their urban function categories. We then use bridge edges to connect nearby nodes across different layers and convert the result into PyTorch Geometric HeteroData format. After that, we build a homogeneous KNN graph, attach degree and engineered features, standardize them, and prepare the final PyG Data object for GraphSAGE training.<\/p>\n
Defining and Training a GraphSAGE Model for POI Classification<\/strong><\/h3>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
perm = torch.randperm(N, generator=torch.Generator().manual_seed(SEED))\nn_tr, n_va = int(0.6 * N), int(0.2 * N)\ntrain_mask = torch.zeros(N, dtype=torch.bool); train_mask[perm[:n_tr]] = True\nval_mask   = torch.zeros(N, dtype=torch.bool); val_mask[perm[n_tr:n_tr + n_va]] = True\ntest_mask  = torch.zeros(N, dtype=torch.bool); test_mask[perm[n_tr + n_va:]] = True\nclass GraphSAGE(torch.nn.Module):\n   def __init__(self, in_dim, hidden, out_dim, p=0.3):\n       super().__init__()\n       self.c1 = SAGEConv(in_dim, hidden)\n       self.c2 = SAGEConv(hidden, hidden)\n       self.lin = torch.nn.Linear(hidden, out_dim)\n       self.p = p\n   def forward(self, x, ei, return_emb=False):\n       h = F.relu(self.c1(x, ei))\n       h = F.dropout(h, p=self.p, training=self.training)\n       h = F.relu(self.c2(h, ei))\n       out = self.lin(h)\n       return (out, h) if return_emb else out\nmodel = GraphSAGE(data.x.shape[1], 64, num_classes)\nopt = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)\ndef evaluate(mask):\n   model.eval()\n   with torch.no_grad():\n       pred = model(data.x, data.edge_index).argmax(1)\n   yt, yp = y[mask].numpy(), pred[mask].numpy()\n   return accuracy_score(yt, yp), f1_score(yt, yp, average=\"macro\")\nprint(\"n--- Training GraphSAGE ---\")\nbest_val, best_state = 0.0, None\nfor epoch in range(1, 201):\n   model.train(); opt.zero_grad()\n   out = model(data.x, data.edge_index)\n   loss = F.cross_entropy(out[train_mask], y[train_mask])\n   loss.backward(); opt.step()\n   if epoch % 20 == 0:\n       va_acc, va_f1 = evaluate(val_mask)\n       if va_acc > best_val:\n           best_val, best_state = va_acc, {k: v.clone() for k, v in model.state_dict().items()}\n       print(f\"epoch {epoch:3d} | loss {loss.item():.3f} | val_acc {va_acc:.3f} | val_f1 {va_f1:.3f}\")\nif best_state: model.load_state_dict(best_state)\nte_acc, te_f1 = evaluate(test_mask)\nprint(f\"nTEST  accuracy={te_acc:.3f}  macro-F1={te_f1:.3f}\")\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We split the graph nodes into training, validation, and test masks so the model can learn and be evaluated properly. We define a two-layer GraphSAGE model that learns node representations from both node features and graph structure. We train the model for 200 epochs, monitor validation accuracy and macro-F1, save the best model state, and finally report test performance.<\/p>\n
Visualizing Embeddings and Running a Heterogeneous GNN Forward Pass<\/strong><\/h3>\n
\n
\n
\n
\n
<\/div>\n
<\/div>\n
<\/div>\n<\/div>\n
Copy Code<\/span>Copied<\/span>Use a different Browser<\/span><\/a><\/div>\n<\/div>\n
model.eval()\nwith torch.no_grad():\n   logits, emb = model(data.x, data.edge_index, return_emb=True)\n   pred = logits.argmax(1).numpy()\nemb2d = PCA(n_components=2).fit_transform(emb.numpy())\nfig, axes = plt.subplots(1, 2, figsize=(15, 6))\nfor cls in range(num_classes):\n   m = y.numpy() == cls\n   axes[0].scatter(emb2d[m, 0], emb2d[m, 1], s=10, label=CLASS_NAMES[cls], alpha=0.7)\naxes[0].set_title(\"GraphSAGE node embeddings (PCA), coloured by TRUE class\")\naxes[0].legend(fontsize=8); axes[0].set_xticks([]); axes[0].set_yticks([])\nplot_gdf = nodes.copy(); plot_gdf[\"pred\"] = pred\nplot_gdf[\"pred_name\"] = [CLASS_NAMES[p] for p in pred]\nplot_gdf.plot(ax=axes[1], column=\"pred_name\", legend=True, markersize=12, cmap=\"tab10\")\naxes[1].set_title(\"Predicted urban function (mapped back to geography)\")\naxes[1].set_axis_off()\ntry:\n   import contextily as ctx\n   ctx.add_basemap(axes[1], crs=plot_gdf.crs, source=ctx.providers.CartoDB.Positron)\nexcept Exception:\n   pass\nplt.tight_layout(); plt.show()\nif hetero is not None:\n   try:\n       for nt in hetero.node_types:\n           hetero[nt].x = hetero[nt].x.float()\n       class HGNN(torch.nn.Module):\n           def __init__(self, hid, out):\n               super().__init__()\n               self.c1 = SAGEConv((-1, -1), hid)\n               self.c2 = SAGEConv((-1, -1), out)\n           def forward(self, x, ei):\n               x = {k: F.relu(v) for k, v in self.c1(x, ei).items()}\n               return self.c2(x, ei)\n       hmodel = to_hetero(HGNN(32, 16), hetero.metadata(), aggr=\"sum\")\n       out_dict = hmodel(hetero.x_dict, hetero.edge_index_dict)\n       print(\"nHeterogeneous GNN output embedding shapes:\")\n       for nt, t in out_dict.items():\n           print(f\"   {nt}: {tuple(t.shape)}\")\n   except Exception as e:\n       print(\"Hetero GNN forward skipped:\", e)\nprint(\"n\"\u2705\" Done \u2014 proximity comparison, hetero construction, and a trained spatial GNN.\")\n<\/code><\/pre>\n<\/div>\n<\/div>\n
We use the trained GraphSAGE model to extract node embeddings and predictions from the homogeneous graph. We reduce the learned embeddings with PCA and visualize them alongside a geographic prediction map to understand how the model separates urban functions. We also run a heterogeneous GNN forward pass with to_hetero, showing that the tutorial supports both homogeneous training and heterogeneous graph experimentation.<\/p>\n
Key Takeaways<\/strong><\/h3>\n