Alleviating neighbor bias: augmenting graph self-supervise learning with structural equivalent positive samples