Because of the extremely low interpretability of unsupervised learning, the most difficulty task is to make sure if the "good" results in some datasets can generalize to other datasets even if we can test it at a high cost in rare situations. It seems that this paper cannot solve this fundamental problem and thus there are still too much uncertainty and risk when we choose method based on the result that looks good.
Because of the extremely low interpretability of unsupervised learning, the most difficulty task is to make sure if the "good" results in some datasets can generalize to other datasets even if we can test it at a high cost in rare situations. It seems that this paper cannot solve this fundamental problem and thus there are still too much uncertainty and risk when we choose method based on the result that looks good.