Practically this almost never works out. The 10% cluster is using a very small dataset and will produce inferior results. If you train a model based on only 10

Practically this almost never works out. The 10% cluster is using a very small dataset and will produce inferior results. If you train a model based on only 10 people, you're prone to overfit that small sample