Unsupervised selection and characterisation of Little Red Dots in JWST surveys with manifold learning

Michele Ginolfi, Filippo Mannucci, Alessandro Marconi, Francesco D'Eugenio, Giacomo Venturi, Francesco Belfiore, Giovanni Cresci, Caterina Bracci, Stefano Carniani, Alessandra Cozzi, Roberto Maiolino, Guido Risaliti

First listed 2026-07-28 | Last updated 2026-07-24

Abstract

Little Red Dots (LRDs) are compact, red sources discovered at high redshift by JWST whose physical nature and selection function remain debated. We investigate whether an unsupervised machine-learning approach applied to multi-band photometry can identify LRD-like objects, and other populations, without relying on predefined colour cuts. Using UMAP, a manifold-learning (dimensionality-reduction) method, we place ~242,000 isolated, well-measured sources from the ASTRODEEP-JWST catalogue on a two-dimensional map, where objects with similar broadband colours, morphology, and photometric redshift lie close together. We then use spectroscopically confirmed LRDs to identify where LRD-like objects lie within this map, compare the resulting areas with published colour cuts, and validate our data-driven selection against archival NIRSpec spectra from the DJA. We find that the spectroscopically selected LRDs concentrate in two well-defined regions with no colour cut imposed, tracing populations that differ mainly in redshift, a difference imprinted in their broadband colours. The main region reaches a purity of ~0.78 at ~0.82 completeness on the spectroscopically classified subset, competitive with, or cleaner than, literature colour cuts, and yields ~100 additional candidates. We also test the method as a general tool for population discovery: the manifold recovers the locations of brown dwarfs and broad-line AGN with no explicit criterion, and isolates rare pathological outliers. Overall, unsupervised manifolds, anchored by sparse high-confidence spectroscopic labels, provide an efficient, assumption-light framework for characterising populations, comparing selection methods on a common basis, and discovering rare objects in large photometric datasets.

Short digest

Ginolfi et al. apply UMAP manifold learning to 242,000 isolated, well-measured ASTRODEEP-JWST sources using eight-band photometry, morphology, and photometric redshifts, then use spectroscopically confirmed LRDs as sparse anchors rather than imposing colour cuts. Confirmed LRDs occupy two compact loci on the manifold, primarily separated by redshift and rest-UV luminosity; the main locus achieves 0.78 purity at 0.82 completeness on the spectroscopic subset and identifies about 100 additional, generally bluer rest-optical candidates. The result provides a common, data-driven framework for comparing inconsistent LRD selections while showing that broadband colours carry most of the LRD signal, and that the same embedding can recover brown dwarfs, broad-line AGN, and photometric pathologies as distinct populations.

Key figures to inspect

• Figure 2. Shows the two manifold-defined LRD anchor loci and their relation to published colour-selected samples, establishing the paper's central selection geometry without imposing an LRD colour cut.
• Figure 3. Provides the quantitative completeness-purity comparison on a common parent sample, including the main and secondary loci and literature selections, making the claimed 0.78-purity and 0.82-completeness performance directly interpretable.
• Figure 5. Demonstrates that the newly selected photometric candidates reproduce the characteristic V-shaped stacked SEDs of spectroscopically confirmed LRDs in both loci, supplying the key population-level validation.
• Figure 10. Identifies broadband colours as the dominant carrier of the LRD signature, while morphology and photometric redshift contribute little on their own, clarifying the physical and methodological basis of the manifold localisation.