Text this: A supervised variational autoencoder framework for dimensionality reduction and predictive modeling in high-dimensional socioeconomic data