I'm analysing an RNA-Seq datasets with edgeR and using calcNormFactors to estimate normalisation factors. In one of my experimental groups all samples have normalisation factors above 1,  while in another group all samples have normalisation factors below 1. Library sizes do not vary systematically between the groups.

What could be an explanation for this pattern?

This is a standard effect of composition bias - see Robinson and Oshlack (2010) for more details. More genes are upregulated (and/or with larger log-fold changes) in one group compared to the other, resulting in greater composition bias whereby the coverage of all other non-DE genes is suppressed during sequencing. This leads to consistently smaller normalization factors for all samples in the first group, to counter the suppressive effect.

Furthermore, note that the returned normalization factors average to 1 across all samples (for ease of interpretation, and to ensure that the effective library sizes are on the same scale as the originals). This naturally means that if the factors for one group are below 1, the factors for the other group must be above 1.

I wouldn't worry about it.