Would it be possible instead to decompose it into a rotation and translation matrix and then average the angle of rotation and translation separately instead and then recombine?
According to https://igl.ethz.ch/projects/skinning/stretchable-twistable-bones/, DQS can help solve the candy wrapper artifact but cannot solve another kind of artifacts called "stretching artifact", which is caused by large scale stretch of body components. The link above proposed a method called stretchable, twistable bones skinning (STBS) to further solve this problem. Moreover the link also contains a visual comparison of different methods.
@briana, great question! The problem is not the translation, but the rotation matrix, this figure could've said SO(3) instead.