Toward Optimal Community Detection: From Trees to General Weighted Networks