UC Berkeley

Context. The growing volume of spectroscopic data in modern surveys motivates the development of data-driven approaches that complement traditional emission-line diagnostics for the identification of active galactic nuclei (AGNs). Aims. We develop and validate a machine-learning (ML) framework for galaxy and AGN classification that exploits the full optical spectrum, using unsupervised learning to capture spectral features within a semi-supervised classification framework that complements classical approaches. Methods. We adapted the SPENDER autoencoder architecture to compress galaxy spectra obtained with the Dark Energy Spectroscopic Instrument (DESI) into a low-dimensional latent space that encodes physically meaningful label-agnostic information. We introduced a quantitative classification method via a k-d tree nearest-neighbor search in latent space. The model was trained on 50 222 spectra from the DESI main survey, corresponding to the internal Guadalupe dataset, released as part of data release 1 (DR1), and restricted to z ≤ 0.5. We validated the performance using labels derived from the FastSpecFit emission line measurements defining seven galaxy classes: AGN, broad-line (BL), composite, star-forming, passive, retired, and other. Results. The model achieves a high performance in identifying AGN and BL sources, with accuracies of 0.952 and 0.965, respectively. The model reliably recovers AGN and BL sources even in spectra with a low signal-to-noise ratio and identifies AGNs missed by single-diagnostic standard methods. Our classification metrics are benchmarked against traditional diagnostics, and we show that they represent lower limits of the true model performance. The latent space learned by the autoencoder not only separates AGN and non-AGN galaxies, but also correlates with key physical properties such as stellar mass and star formation rate, demonstrating that the method captures astrophysical meaningful features. Conclusions. This work demonstrates that unsupervised spectral representation learning implemented within a semi-supervised classification framework offers a powerful and scalable pathway toward more complete and reliable AGN catalogs for current and upcoming large spectroscopic surveys.