Partitioning clustering is generally performed using K-modes cluster algorithms, which work well for large datasets. A Kmodes technique involve random chosen initial cluster centre (modes) as seed, which lead toward that problem clustering results be regularly reliant on the choice initial cluster centre and non-repeatable cluster structure may be obtain. K-Modes technique has been widely applied to categorical data a clustering in replace means through modes. The pervious algorithms select the attributes on frequency basis but not provided better result. Proposed algorithm select attributes on information gain basis which provide better result. Experimental results showing the proposed technique provided better accuracy.