PyramidDiff (WPP 2024)

A small conditional diffusion model (1D UNet, ~7.3M parameters) of single-age, sex-specific population structure. A pyramid is a (2 x 101) array of population shares, ages 0-100+.

Training run

quick run: 3000 optimisation steps.

What it can do

  • Generate plausible pyramids from summary conditions (median age, % under 15, % 65+, sex ratio, year), with sampling-based uncertainty.
  • Repair age heaping (digit preference at 0/5): start from a classical window-total-preserving smoothing (PCHIP), partially noise it, and denoise with the model while re-imposing the observed totals of 5-year windows (centred on the heaping digits) at every denoising step. The model therefore only decides the within-window shape.
  • Inpaint missing ages (evaluated for one hidden 20-year span, below).
  • Flag implausible pyramids via denoising error.

Training data

UN World Population Prospects 2024 (Medium variant), single-age population by sex, countries/areas, 1950-2023. WPP pyramids are modelled estimates, not raw census counts. Split by country: 80/10/10 train/validation/test. Projections for 2024-2100 were also used (training countries only); validation and test are purely historical. Source: United Nations, Department of Economic and Social Affairs, Population Division (2024). World Population Prospects 2024. https://population.un.org/wpp/ Check the UN terms of use for the data before reuse.

Evaluation: age-heaping repair (held-out test countries, 1702 pyramids; heaping is SIMULATED)

Mean absolute error of single-age shares, per mille of total population (lower is better):

Method MAE
Heaped input (no repair) 1.687
5-point moving average + window projection (classical) 0.128
PCHIP window-total interpolation (classical) 0.102
PyramidDiff, raw-input start, one final projection 0.119
PyramidDiff, raw-input start, per-step projection 0.099
PyramidDiff, PCHIP start, per-step projection (default) 0.096

On held-out countries the default PyramidDiff variant improves on the best classical baseline (0.096 vs 0.102 per mille). Mean Whipple index: truth 100, heaped 193, PCHIP 100, PyramidDiff default 100. Anomaly score AUROC (clean vs simulated-heaped): 1.000.

Evaluation: inpainting (held-out test countries, 1702 pyramids, ages 30-49 hidden in both sexes)

MAE on the hidden ages only, per mille of total population:

Method MAE
Hidden ages filled with the mean share 1.486
Linear interpolation of log-share across the gap 0.341
PyramidDiff inpainting 0.752

PyramidDiff does not improve on linear interpolation for this hidden span. Other spans, and the conditioning used when ages are missing, were not tested.

Limitations

  • Heaping was evaluated against a synthetic heaping process; real survey heaping differs (and varies by country, sex and age).
  • Trained on smooth modelled estimates, so it will under-represent real-world irregularities (migration shocks, war, small populations).
  • Repair assumes heaping moves people by at most 2 years; it cannot fix large age misreporting, or sex-specific selective omission.
  • Research prototype. Not for official statistics or individual-level decisions.

Usage

from huggingface_hub import snapshot_download
import importlib.util, os
p = snapshot_download("SallySims/pyramiddiff-wpp2024")
spec = importlib.util.spec_from_file_location("pyramiddiff", os.path.join(p, "pyramiddiff.py"))
P = importlib.util.module_from_spec(spec); spec.loader.exec_module(P)
model, diff, norm = P.load_pretrained(p)
samples = P.generate(model, diff, norm, median_age=20, u15=0.42, p65=0.03, sex_ratio=99, year=2005)
Downloads last month
-
Safetensors
Model size
7.28M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Dataset used to train SallySims/pyramiddiff-wpp2024