Capturing Non‐Linear Neighborhood Structure: An Autoencoder Approach to Census‐Based Dimensionality Reduction for Residential Differentiation

Alex Singleton; Owen Goodwin; Stef De Sabbata (2026). Geographical Analysis, 58(4). DOI: 10.1111/gean.70059

Abstract

Dimensionality reduction is fundamental to applied spatial data analysis, condensing high‐dimensional indicators into parsimonious representations for mapping and modeling. Principal component analysis (PCA) remains a dominant approach, yet its linearity assumptions constrain its capacity to capture the complex, non‐linear dependencies characteristic of spatial socio‐economic data. We evaluate the effectiveness of autoencoders as a flexible, parametric alternative for constructing composite measures of neighborhood structure. Through encoder‐decoder architectures, AEs learn non‐linear representations that better reflect underlying data complexity, enable projection of new observations without refitting, and provide reconstruction‐based diagnostics that enhance transparency. We demonstrate the approach using small‐area data from the 2021 Census of England and Wales, evaluating how effectively AEs summarize spatial variability and recover composite dimensions of residential differentiation compared with PCA.

Extended Summary

Can neural networks capture aspects of neighbourhood structure that principal component analysis misses? This paper evaluates autoencoders as a non-linear alternative to principal component analysis (PCA) for building composite measures of residential differentiation from small-area census data.

Composite indicators, single numerical summaries that condense dozens or hundreds of census variables into a handful of dimensions, underpin much applied spatial analysis, from deprivation indices to geodemographic classifications. PCA has long dominated this work, but it assumes the structure of the data can be described by linear combinations of the input variables. Socio-economic and demographic data rarely behave that way, containing threshold effects, interactions, and relationships that vary across space. Non-parametric manifold learning methods such as t-SNE and UMAP relax the linearity assumption, but they preserve local rather than global structure, offer no reliable way to reconstruct the original variables, require sensitive parameter choices, and typically cannot project new areas without refitting the whole model. Those constraints matter for indicators embedded in policy workflows, which must be updatable, transparent, and consistent over time.

The research uses 408 variables from the 2021 Census of England and Wales, extracted in bulk from Nomis for all 188,880 Output Areas, the smallest census geography and typically 40 to 250 households, covering demography, migration, ethnicity, identity, language, religion, work, travel, housing, education, and health. A symmetric encoder-decoder network with three dimensionality-reducing layers and Leaky ReLU activations was trained using a normalised mean squared error loss over 500 epochs, with mini-batches of 1% of observations and ten random initialisations per configuration. Latent dimensionalities from 2 to 128 were benchmarked against PCA decompositions using the equivalent number of components, with reconstruction error measured as root mean squared error in percentage-point units. A spatially blocked cross-validation, withholding entire Local Authority Districts across five folds, tested whether any gains simply reflected overfitting.

The autoencoder reconstructed the census data more accurately than PCA at every dimensionality tested, with the largest advantage where compression was greatest. At 100 dimensions, a typical operational specification, mean reconstruction error fell by 19.4%, and the advantage held on the spatially independent hold-out districts. Autoencoder residuals also showed consistently lower spatial autocorrelation, indicating that more of the geographically structured variance had been captured. Reconstruction errors were higher for both methods in deprived and densely populated neighbourhoods, but the autoencoder narrowed that gap, improving most in deprivation deciles 1 to 4, the densest areas, and the Output Area Classification Supergroups covering multicultural, student, and transient populations. It also handled geographically concentrated phenomena, such as Cornish and Welsh national identity, that PCA represents poorly. Applying k-means clustering with eight clusters to the latent space produced a spatially coherent geodemographic classification corresponding closely to the 2021 Output Area Classification, although around one Output Area in six received a different label than under PCA. Decoding cluster centroids back into census variables matched observed cluster means to within 0.3 percentage points.

The study demonstrates that the choice of dimensionality reduction is not neutral for policy applications such as service targeting and resource allocation. Autoencoders offer out-of-sample projection and reconstruction-based diagnostics, while remaining less directly interpretable than PCA loadings and more demanding computationally.

Key Findings

• Autoencoders achieved lower reconstruction error than PCA at every latent dimensionality from 2 to 128, improving by 19.4% at 100 dimensions
• Gains persisted under spatially blocked cross-validation withholding whole Local Authority Districts, confirming more efficient encoding rather than overfitting of training areas
• Autoencoder residuals showed consistently lower spatial autocorrelation than PCA, capturing more of the geographically structured variance in small-area census data
• Improvements were largest in deprived, densely populated, and multicultural neighbourhoods, reducing uneven information loss across socio-economic and spatial gradients
• Decoding k-means cluster centroids back to census variables matched observed cluster means within 0.3 percentage points, supporting policy-legible interpretation

Citation

PDF Download BibTeX

@article{singleton2026capturing,
  author = {Alex Singleton; Owen Goodwin; Stef De Sabbata},
  title = {Capturing Non‐Linear Neighborhood Structure: An Autoencoder Approach to Census‐Based Dimensionality Reduction for Residential Differentiation},
  journal = {Geographical Analysis},
  year = {2026},
  volume = {58(4)},
  doi = {10.1111/gean.70059}
}