A novel data-driven approach to extract stellar population properties from galaxy spectra using absorption indices
In an era of highly complex machine learning methods that often are informative but not straightforward to interpret, Principal Component Analysis (PCA) offers a simple, easily interpretable approach. With no fitting parameters, it extracts the most salient statistical trends in data without the need for training sets. In this paper, we explore a large range of composite stellar population models defined for detailed analyses of galaxy spectra from surveys. Six of the most prominent spectral indices are targeted to visualize a PCA-based latent space created by the model data. The age-metallicity degeneracy is broken in the 3-dimensional space spanned by the first three eigenvectors, but we emphasize that non-trivial combinations of all six absorption indices are needed for this. Moreover, the last eigenvector suggests an intriguing tug of war between two Balmer indices: H$γ_A$ and $Hδ_A$, that can help discern the presence of recent bursting behaviour, as it exploits the different behaviour of the two indices over timescales $\sim$0.5-1 Gyr. Comparisons can be made between SDSS and LEGA-C galaxy spectra based on the latent space created by the models. This method, based on pure data, produces excellent results in agreement with standard SPS model fitting techniques, allowing for the study of stellar populations in a variety of surveys or observational/synthetic databases on solid ground.