Evolutionary　K-Means　(EKM)　is　a　non-parametric　approach　proposed　to　improve　K-Means　algorithm.　Current　EKM　approaches　are　ineffective　in　deciding　the　correct　cluster　number　of　real　datasets.　This　paper　uses　instance-level　constraints　to　solve　this　problem　and　presents　a　Constrained　Silhouette　(CS)　based　algorithm,　namely　CS-EAC.　Firstly　CS　is　defined　to　combine　constraints　into　the　computation　of　Silhouette　Information　(SI).　Updated　from　the　Fast　Evolutionary　Algorithm　for　Clustering　algorithm　(F-EAC),　CS-EAC　uses　CS　instead　of　SI　to　guide　the　genetic　operations.　Experimental　results　suggest　that　CS-EAC　is　effective　in　both　deciding　the　correct　number　of　clusters　and　improving　the　accuracy　of　clustering　for　real　datasets.