Is $mu$ the renormalization or factorization scale in the DGLAP equations?

The DGLAP equations read
$$frac{partial f_i(x,mu^2)}{partiallnmu^2}=sum_jint^1_xfrac{dz}{z}P_{ij}(z,alpha_s(mu^2))f_jleft(frac{x}{z},mu^2right),$$
where the $f_i$ are the parton distribution functions (PDFs), $P_{ij}$ are the so-called splitting kernels and $x,z$ are longitudinal momentum fractions.

But what is $mu$? In this paper on p.26 John Collins says it is the renormalisation scale, which enters the PDF via dimensional regularization. But I have already seen other authors claim that it is a factorization scale, e.g. here on p.2.

So, which one is it?