When unlearning is free: leveraging low influence points to reduce computational costs