An efficient projected partition algorithm based on the order among genes

Most existing methods perform the projected partition over gene expression data based on the untrue assumption of independence among genes. To address the problem, we propose two novel projected partition algorithms, PPA and PPA+. The basic idea of PPA is to take the order among genes as the criterion of phenotype structure discovery. Specially, in PPA, no any specific data distribution assumption is needed. By transforming the expression values into sequential data, PPA employs the branch and bound paradigm to conduct an efficient depth-first traverse over the sample enumeration space, where a user-specific favorite function is developed. Further, based on the hill-climbing strategy and the devised evaluation function, two different strategies are proposed to make the results satisfy different customer-requirements. Efficient pruning and optimization strategies are also devised to further improve the performance of the algorithms. We conducted the performance comparison of PPA and some related alternatives on five real Microarray datasets. The results show that PPA is more effective and efficient. More important, PPA may provide a fire-new insight into the pathogenesis problem.