Progressive Trust Propagation for Urban Region Classification

Code for reproducing the experiments in "When Labels Are Far Away: Progressive Trust Propagation for Urban Region Classification under Spatial Label Scarcity."

Repository Structure

├── config.py               # Global constants (cities, training hyperparameters, device)
├── data_utils.py           # Data loading and preprocessing (normalization, mask creation)
├── base_models.py          # All base classifiers: RF, MLP, GCN, GAT, GraphSAGE, SGC, APPNP, GPRGNN, H2GCN
├── trainer.py              # Training loop (early stopping) and Random Forest wrapper
├── evaluation.py           # Metric computation (AUC, Recall, Macro-F1, Micro-F1)
├── post_processing.py      # Post-processing methods: LP, C&S, and PTP (our method)
├── run_experiment.py        # Main experiment runner (two-phase: cache predictions → evaluate)
└── Graph_Dataset_Block_Balanced/   # Data directory (expected location for .pt graph files)

Requirements

• Python ≥ 3.9
• PyTorch ≥ 2.0
• PyTorch Geometric
• scikit-learn
• pandas, numpy

Data

Each city benchmark is stored as a single PyG Data object (e.g., London_05pct.pt) containing node features, edge indices, labels, train/test masks, and the Borderland (bridge) mask. Place these files under Graph_Dataset_Block_Balanced/.

Note: The dataset files are too large to host on GitHub. They will be shared via Google Drive.

Quick Start

Run a single city with PTP only (1 run):

python run_experiment.py --city London --ratio 05pct --n_runs 1 --imlp_only

Full reproduction (all cities, 10 runs, all post-processing methods):

python run_experiment.py --city all --ratio 05pct --n_runs 10

Two-Phase Execution

The experiment runner separates training from evaluation. This avoids retraining base models when sweeping post-processing configurations.

Phase 1 — Train base models and cache predictions:

python run_experiment.py --city all --phase cache

Predictions are saved as .pt files in the cache/ directory.

Phase 2 — Apply post-processing and evaluate:

python run_experiment.py --city all --phase eval

By default, --phase both runs both phases sequentially.

Command-Line Options

Argument	Default	Description
--city	all	City to run: Harbin, London, Paris, or all
--ratio	05pct	Training ratio: 05pct through 25pct, or all
--n_runs	10	Number of runs (seeds 42–51)
--base_models	all 9	Subset of base models, e.g., --base_models MLP GCN
--phase	both	cache (train only), eval (post-process only), or both
--base_only	off	Skip all post-processing
--imlp_only	off	Run only base + PTP (skips LP and C&S)
--cache_dir	cache	Directory for cached base predictions
--output_dir	results	Directory for result CSVs

Outputs

Results are written to results/:

• results_{City}_{Ratio}.csv — Per-run metrics for every (base model, post-processing, region) combination.
• all_results.csv — Combined results across all cities and ratios.
• summary.csv — Aggregated mean ± std for all metrics.

PTP Hyperparameters

PTP uses fixed hyperparameters across all experiments (set in run_experiment.py):

Symbol	Parameter	Value
K	num_stages	3
α	alpha_calibrate / alpha_aggregate	0.999
γ	gamma	0.5
τ	conf_threshold	0.8
λ	dense_weight	0.8
β	sharpness	10.0
τ_m	margin_thresh	0.3
T	num_iter_calibrate / num_iter_aggregate	200

To modify these, edit the IMLPConfig(...) instantiation in run_experiment.py.