Shallow shotgun metagenomics has been recently suggested as a promising strategy to study human microbiota, providing nearly identical taxonomic profiles than deep shotgun metagenomics with a sequencing cost similar to metabarcoding. With shallow sequencing approach (typically <1M reads/samples), taxonomic profiles are directly built by mapping reads on a catalog of reference genomes, without assembly step. In the present study, we first used simulated data set to design a dedicated workflow in order to obtain reliable taxonomic profiles from shallow sequencing reads. We propose a novel data-driven filtering method based on machine learning techniques that largely outperformed basic filtering methods. We then used this approach on 3 real data sets, covering patients from several continents and clinical conditions. Even if one looses some information like rare taxa, our results clearly show that shallow shotgun metagenomics is able to correctly retrieve structures like differences between groups of patients and diagnosis-like classification.