Unsupervised selection and characterisation of little red dots in JWST surveys with manifold learning

Little red dots (LRDs) are compact, red sources discovered at high redshifts by JWST, and their physical nature and selection function remain a topic of debate. Here, we investigate whether an unsupervised machine-learning approach applied to multi-band photometry has the capacity to identify LRD-like objects, along with other populations, without relying on predefined colour-cuts. Using UMAP, a manifold-learning (dimensionality-reduction) method, we placed ∼242,000 isolated, well-measured sources from the ASTRODEEP-JWST catalogue on a 2D map, where objects with similar broadband colours, morphology, and photometric redshift lie close together. We then used spectroscopically confirmed LRDs to identify where LRD-like objects lie within this map. We compared the resulting areas with published colour cuts and validated our data-driven selection against archival NIRSpec spectra from the DJA. We find that the spectroscopically selected LRDs are concentrated in two well-defined regions with no colour-cut imposed, tracing populations that differ mainly in redshift; a difference that is imprinted on their broadband colours. The main region reaches a purity of ≃0.78 at ≃0.82 completeness on the spectroscopically classified subset. This result is competitive with (or cleaner than) literature colour-cuts, and yields ∼100 additional candidates. We also tested the method as a general tool for population discovery: the manifold recovers the locations of brown dwarfs and broad-line active galactic nuclei 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.