ICE-ID Dataset

Overview

ICE-ID is a benchmark dataset of Icelandic census records (1703–1920) for longitudinal identity resolution. It includes cleaned tabular features and a temporal graph of person records across census waves.

Files

raw_data/
├─ people.csv
├─ manntol_einstaklingar_new.csv
├─ parishes.csv, districts.csv, counties.csv

artifacts/
├─ row_labels.csv           # row_id, person mapping
├─ rows_with_person.csv     # linked subset of rows
├─ iceid_ml_ready.npz       # CSR matrix with numeric, one-hot, ordinal features
├─ temporal_graph.pt        # PyTorch Geometric data (edge_index, node_id)

Requirements

• Python 3.7+
• numpy, pandas, scipy, scikit-learn, torch, torch-geometric

Install via:

pip install numpy pandas scipy scikit-learn torch torch-geometric

Preprocessing

Run the preprocessing script to generate artifacts:

python preprocess.py --data-dir raw_data --out-dir artifacts

Usage

• Load iceid_ml_ready.npz for ML features (rows × features).
• Use row_labels.csv to map rows to persons.
• Load temporal_graph.pt in PyG:

  import torch
  from torch_geometric.data import Data
  d = torch.load('artifacts/temporal_graph.pt')
  graph = Data(edge_index=d['edge_index'])
  graph.node_id = d['node_id']
  

Code requirements

• The repository has been split across Huggingface and GitHub. As such, the code required to run the program and interact with the can be found at https://github.com/IIIM-IS/ICE-ID-2.0. Please, match the folder structure as indicated in the README found therein.

---

license: mit