Symbiotic binaries in the Gaia data. II. Symbiotic candidates from general variability classification in DR3

We investigate the reliability of the symbiotic star class in the general variability classification of Gaia DR3 and search for new genuine symbiotic systems among the objects assigned to this category. The sample contains 649 sources, including 246 previously known symbiotic stars, 61 literature candidates, and 339 new candidates proposed by the Gaia variability pipeline. Diagnostics based on the Gaia colour-magnitude diagram, near-infrared photometry, and the pseudo-equivalent width of H$\alpha$ indicate that a large fraction of the new candidates are likely contaminants, predominantly pulsating red giants. To quantify the contamination, we constructed a Random Forest classifier trained on confirmed symbiotic stars and on Mira and semi-regular variables, using Gaia photometry, variability parameters, H$\alpha$ measurements from XP spectra, and infrared colours. The classifier reaches a balanced accuracy of $\approx$0.94 and efficiently separates most symbiotic binaries from single evolved stars, leaving only eight strong candidates among the 339 newly proposed objects. Follow-up spectroscopy confirms three new symbiotic stars through the presence of high-excitation emission lines, while several additional objects remain possible symbiotics. Our results show that the Gaia DR3 variability classification efficiently recovers known symbiotic stars but has low purity due to overlap with pulsating red giants. The small number of newly confirmed systems implies that the discrepancy between predicted and observed Galactic symbiotic populations remains unresolved, although Gaia provides a powerful basis for future searches combining variability, spectroscopic indicators, and multi-wavelength data.