3dSAGER: Geospatial Entity Resolution over 3D Objects

3dSAGER (3D Spatial-Aware Geospatial Entity Resolution) is an end-to-end pipeline for geospatial entity resolution over 3D objects. Unlike traditional methods that rely on spatial proximity, textual metadata, or external identifiers, 3dSAGER captures intrinsic geometric characteristics to robustly match spatial objects across datasets, even when coordinate systems are incompatible.

A key component of 3dSAGER is BKAFI, a lightweight and interpretable blocking method that efficiently generates high-recall candidate sets.

📄 The paper is currently under review for SIGMOD 2026.

---

📦 Dataset

You can download the The Hague dataset using the following link:

🔗 Download dataset

We provide further instructions for working with the The Hague dataset in the Working_with_The_Hague_Dataset.ipynb notebook.

---

🚀 Running Experiments

You can run the experiments using the provided shell script:

bash run_experiments.sh

In the run_experiments.sh script, you must define the evaluation mode by setting eval_mode to either "blocking" or "matching". For a single basic experimental configuration, we recommend using the "small" or "large" dataset size, setting bkafi_criterion="feature_importance", normalizations=True, and sdr_factor=False.

Here are some of the main variables to be configured in config.py:

Variable	Description
dataset_name	Dataset key: "Hague"
evaluation_mode	"blocking" or "matching"
dataset_size_version	Size variant: "small", "large"
matching_cands_generation	How to generate candidate pairs ("blocking-based" or "negative_sampling"). Select "blocking-based"
neg_samples_num	Number of negative samples per positive. Select 2.
seeds_num	Number of seeds for experiments. Select 3.
blocking_method	'bkafi'
cand_pairs_per_item_list	List of candidate counts per object
nn_param	Number of nearest neighbors to retrieve. Select 20.
sdr_factor	False
bkafi_criterion	Feature selection strategy: 'feature_importance' or 'std'. Select 'feature_importance'
model_to_use	Default classifier used for prediction
model_list	Available models for matching
blocking_model	Classifier used in the blocking stage
params_dict	Hyperparameter grid per model (for cross-validation)